agora inbox for [email protected]
help / color / mirror / Atom feed[PATCH v21 3/3] Add Assert guard to detect permission check before lock regressions
188+ messages / 2 participants
[nested] [flat]
* [PATCH v21 3/3] Add Assert guard to detect permission check before lock regressions
@ 2026-04-27 14:23 Bertrand Drouvot <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Bertrand Drouvot @ 2026-04-27 14:23 UTC (permalink / raw)
Add instrumentation under USE_ASSERT_CHECKING to detect cases where
object_aclcheck() is called on a referenced object before a lock is held on it,
which would widen the TOCTOU window between the permission check and the dependency
recording.
In recordMultipleDependencies() and changeDependencyFor(), for each referenced
object that had a permission check earlier in the same statement, assert that a
lock is already held. This catches any future code that adds a permission check
on a referenced object without first acquiring a lock.
The tracking records each (classId, objectId, mode) from object_aclcheck() and
pg_class_aclcheck(), filtering to only dependency-relevant ACL modes (ACL_CREATE,
ACL_USAGE on non namespaces, ACL_EXECUTE, ACL_TRIGGER, ACL_REFERENCES). Tracking
resets at each ProcessUtility() call. The tracking array is capped at 1024 entries
per statement, which is sufficient for any practical DDL command.
Author: Bertrand Drouvot <[email protected]>
Reviewed-by:
Discussion: https://postgr.es/m/[email protected]
---
src/backend/catalog/aclchk.c | 32 ++++++++++++
src/backend/catalog/pg_depend.c | 52 ++++++++++++++++++++
src/backend/tcop/utility.c | 3 ++
src/include/catalog/aclcheck_track.h | 73 ++++++++++++++++++++++++++++
src/tools/pgindent/typedefs.list | 1 +
5 files changed, 161 insertions(+)
51.5% src/backend/catalog/
46.7% src/include/catalog/
diff --git a/src/backend/catalog/aclchk.c b/src/backend/catalog/aclchk.c
index 007ede997c5..d13667f38a8 100644
--- a/src/backend/catalog/aclchk.c
+++ b/src/backend/catalog/aclchk.c
@@ -76,6 +76,7 @@
#include "parser/parse_type.h"
#include "storage/lmgr.h"
#include "utils/acl.h"
+#include "catalog/aclcheck_track.h"
#include "utils/aclchk_internal.h"
#include "utils/builtins.h"
#include "utils/fmgroids.h"
@@ -3891,6 +3892,8 @@ object_aclcheck_ext(Oid classid, Oid objectid,
Oid roleid, AclMode mode,
bool *is_missing)
{
+ aclcheck_track_record(classid, objectid, mode);
+
if (object_aclmask_ext(classid, objectid, roleid, mode, ACLMASK_ANY,
is_missing) != 0)
return ACLCHECK_OK;
@@ -4093,6 +4096,8 @@ AclResult
pg_class_aclcheck_ext(Oid table_oid, Oid roleid,
AclMode mode, bool *is_missing)
{
+ aclcheck_track_record(RelationRelationId, table_oid, mode);
+
if (pg_class_aclmask_ext(table_oid, roleid, mode,
ACLMASK_ANY, is_missing) != 0)
return ACLCHECK_OK;
@@ -5036,3 +5041,30 @@ RemoveRoleFromInitPriv(Oid roleid, Oid classid, Oid objid, int32 objsubid)
table_close(rel, RowExclusiveLock);
}
+
+/*
+ * Instrumentation to detect permission check before lock regressions.
+ * Only used in assert-enabled builds.
+ */
+#ifdef USE_ASSERT_CHECKING
+AclCheckEntry aclcheck_tracked[ACLCHECK_TRACK_MAX];
+int aclcheck_tracked_count = 0;
+
+void
+aclcheck_track_reset(void)
+{
+ aclcheck_tracked_count = 0;
+}
+
+bool
+aclcheck_track_was_checked(Oid classId, Oid objectId)
+{
+ for (int i = 0; i < aclcheck_tracked_count; i++)
+ {
+ if (aclcheck_tracked[i].classId == classId &&
+ aclcheck_tracked[i].objectId == objectId)
+ return true;
+ }
+ return false;
+}
+#endif /* USE_ASSERT_CHECKING */
diff --git a/src/backend/catalog/pg_depend.c b/src/backend/catalog/pg_depend.c
index 5618e3d26fa..28fa99b2d42 100644
--- a/src/backend/catalog/pg_depend.c
+++ b/src/backend/catalog/pg_depend.c
@@ -18,6 +18,7 @@
#include "access/htup_details.h"
#include "access/table.h"
#include "catalog/catalog.h"
+#include "catalog/aclcheck_track.h"
#include "catalog/dependency.h"
#include "catalog/indexing.h"
#include "catalog/pg_constraint.h"
@@ -27,6 +28,8 @@
#include "catalog/partition.h"
#include "commands/extension.h"
#include "miscadmin.h"
+#include "storage/lmgr.h"
+#include "storage/lock.h"
#include "utils/fmgroids.h"
#include "utils/lsyscache.h"
#include "utils/rel.h"
@@ -107,6 +110,31 @@ recordMultipleDependencies(const ObjectAddress *depender,
if (isObjectPinned(referenced))
continue;
+#ifdef USE_ASSERT_CHECKING
+ /*
+ * If this referenced object had a permission check earlier in this
+ * statement, assert that a lock is already held on it. This ensures
+ * callers acquired the lock before calling object_aclcheck(), not
+ * after. The latter would widen the TOCTOU window between the
+ * permission check and the dependency recording.
+ */
+ if (aclcheck_track_was_checked(referenced->classId, referenced->objectId))
+ {
+ if (referenced->classId == RelationRelationId)
+ Assert(CheckRelationOidLockedByMe(referenced->objectId,
+ AccessShareLock, true));
+ else
+ {
+ LOCKTAG tag;
+
+ SET_LOCKTAG_OBJECT(tag, MyDatabaseId,
+ referenced->classId,
+ referenced->objectId, 0);
+ Assert(LockHeldByMe(&tag, AccessShareLock, true));
+ }
+ }
+#endif
+
/*
* Acquire a lock and check object still exists while recording the
* dependency.
@@ -511,6 +539,30 @@ changeDependencyFor(Oid classId, Oid objectId,
return 1;
}
+#ifdef USE_ASSERT_CHECKING
+ /*
+ * If this referenced object had a permission check earlier in this
+ * statement, assert that a lock is already held on it. This ensures
+ * callers acquired the lock before calling object_aclcheck(), not after.
+ * The latter would widen the TOCTOU window between the permission check and
+ * the dependency recording.
+ */
+ if (aclcheck_track_was_checked(objAddr.classId, objAddr.objectId))
+ {
+ if (objAddr.classId == RelationRelationId)
+ Assert(CheckRelationOidLockedByMe(objAddr.objectId,
+ AccessShareLock, true));
+ else
+ {
+ LOCKTAG tag;
+
+ SET_LOCKTAG_OBJECT(tag, MyDatabaseId,
+ objAddr.classId,
+ objAddr.objectId, 0);
+ Assert(LockHeldByMe(&tag, AccessShareLock, true));
+ }
+ }
+#endif
/*
* Acquire a lock and check object still exists while changing the
* dependency.
diff --git a/src/backend/tcop/utility.c b/src/backend/tcop/utility.c
index 73a56f1df1d..0c4d382bd37 100644
--- a/src/backend/tcop/utility.c
+++ b/src/backend/tcop/utility.c
@@ -21,6 +21,7 @@
#include "access/xact.h"
#include "access/xlog.h"
#include "catalog/namespace.h"
+#include "catalog/aclcheck_track.h"
#include "catalog/pg_authid.h"
#include "catalog/pg_inherits.h"
#include "catalog/toasting.h"
@@ -515,6 +516,8 @@ ProcessUtility(PlannedStmt *pstmt,
Assert(queryString != NULL); /* required as of 8.4 */
Assert(qc == NULL || qc->commandTag == CMDTAG_UNKNOWN);
+ aclcheck_track_reset();
+
/*
* We provide a function hook variable that lets loadable plugins get
* control when ProcessUtility is called. Such a plugin would normally
diff --git a/src/include/catalog/aclcheck_track.h b/src/include/catalog/aclcheck_track.h
new file mode 100644
index 00000000000..5825d7398d3
--- /dev/null
+++ b/src/include/catalog/aclcheck_track.h
@@ -0,0 +1,73 @@
+/*-------------------------------------------------------------------------
+ *
+ * aclcheck_track.h
+ * Instrumentation to detect permission check before lock via Assert in
+ * recordMultipleDependencies() and changeDependencyFor().
+ *
+ * Only active in USE_ASSERT_CHECKING builds.
+ *
+ * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
+ * Portions Copyright (c) 1994, Regents of the University of California
+ *
+ * src/include/catalog/aclcheck_track.h
+ *
+ *-------------------------------------------------------------------------
+ */
+#ifndef ACLCHECK_TRACK_H
+#define ACLCHECK_TRACK_H
+
+#ifdef USE_ASSERT_CHECKING
+
+#include "catalog/pg_namespace.h"
+
+#define ACLCHECK_TRACK_MAX 1024
+
+typedef struct AclCheckEntry
+{
+ Oid classId;
+ Oid objectId;
+} AclCheckEntry;
+
+extern AclCheckEntry aclcheck_tracked[];
+extern int aclcheck_tracked_count;
+
+extern void aclcheck_track_reset(void);
+extern bool aclcheck_track_was_checked(Oid classId, Oid objectId);
+
+/*
+ * Only record aclchecks that are dependency-relevant:
+ * - ACL_CREATE on any object (creating something in a container)
+ * - ACL_USAGE on non-namespace objects (using a type, language, server)
+ * - ACL_EXECUTE on functions
+ * - ACL_TRIGGER, ACL_REFERENCES on relations
+ *
+ * Skip ACL_USAGE on namespaces — that's name resolution, not dependency.
+ */
+static inline void
+aclcheck_track_record(Oid classId, Oid objectId, AclMode mode)
+{
+ /* Skip ACL_USAGE on namespaces (name resolution, not dependency) */
+ if (classId == NamespaceRelationId && !(mode & ACL_CREATE))
+ return;
+
+ /* Only track dependency-relevant modes */
+ if (!(mode & (ACL_CREATE | ACL_USAGE | ACL_EXECUTE | ACL_TRIGGER | ACL_REFERENCES)))
+ return;
+
+ if (aclcheck_tracked_count < ACLCHECK_TRACK_MAX)
+ {
+ aclcheck_tracked[aclcheck_tracked_count].classId = classId;
+ aclcheck_tracked[aclcheck_tracked_count].objectId = objectId;
+ aclcheck_tracked_count++;
+ }
+}
+
+#else /* !USE_ASSERT_CHECKING */
+
+#define aclcheck_track_reset() ((void) 0)
+#define aclcheck_track_record(c, o, m) ((void) 0)
+#define aclcheck_track_was_checked(c, o) (false)
+
+#endif /* USE_ASSERT_CHECKING */
+
+#endif /* ACLCHECK_TRACK_H */
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 8cf40c87043..cb0cc25bda8 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -19,6 +19,7 @@ AbsoluteTime
AccessMethodInfo
AccessPriv
Acl
+AclCheckEntry
AclItem
AclMaskHow
AclMode
--
2.34.1
--FjYZKkuBdH+F20sI--
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
* [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics
@ 2026-07-09 20:21 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 188+ messages in thread
From: Nathan Bossart @ 2026-07-09 20:21 UTC (permalink / raw)
---
src/backend/access/heap/heapam_handler.c | 2 +-
src/backend/access/table/tableam.c | 59 +++++++++++-------------
src/include/access/relscan.h | 8 ++--
3 files changed, 31 insertions(+), 38 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index bf87430cf01..0f24a132564 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -1965,7 +1965,7 @@ heapam_scan_get_blocks_done(HeapScanDesc hscan)
if (hscan->rs_base.rs_parallel != NULL)
{
bpscan = (ParallelBlockTableScanDesc) hscan->rs_base.rs_parallel;
- startblock = bpscan->phs_startblock;
+ startblock = pg_atomic_read_u32(&bpscan->phs_startblock);
}
else
startblock = hscan->rs_startblock;
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index 68ff0966f1c..f2038ea9205 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -421,9 +421,8 @@ table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
bpscan->base.phs_syncscan = synchronize_seqscans &&
!RelationUsesLocalBuffers(rel) &&
bpscan->phs_nblocks > NBuffers / 4;
- SpinLockInit(&bpscan->phs_mutex);
- bpscan->phs_startblock = InvalidBlockNumber;
- bpscan->phs_numblock = InvalidBlockNumber;
+ pg_atomic_init_u32(&bpscan->phs_startblock, InvalidBlockNumber);
+ pg_atomic_init_u32(&bpscan->phs_numblock, InvalidBlockNumber);
pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
return sizeof(ParallelBlockTableScanDescData);
@@ -459,25 +458,22 @@ table_block_parallelscan_startblock_init(Relation rel,
StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
"pg_nextpower2_32 may be too small for non-standard BlockNumber width");
- BlockNumber sync_startpage = InvalidBlockNumber;
BlockNumber scan_nblocks;
/* Reset the state we use for controlling allocation size. */
memset(pbscanwork, 0, sizeof(*pbscanwork));
-retry:
- /* Grab the spinlock. */
- SpinLockAcquire(&pbscan->phs_mutex);
-
/*
* When the caller specified a limit on the number of blocks to scan, set
* that in the ParallelBlockTableScanDesc, if it's not been done by
* another worker already.
*/
- if (numblocks != InvalidBlockNumber &&
- pbscan->phs_numblock == InvalidBlockNumber)
+ if (numblocks != InvalidBlockNumber)
{
- pbscan->phs_numblock = numblocks;
+ uint32 expected = InvalidBlockNumber;
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_numblock, &expected,
+ numblocks);
}
/*
@@ -485,36 +481,35 @@ retry:
* so now. If a startblock was specified, start there, otherwise if this
* is not a synchronized scan, we just start at block 0, but if it is a
* synchronized scan, we must get the starting position from the
- * synchronized scan machinery. We can't hold the spinlock while doing
- * that, though, so release the spinlock, get the information we need, and
- * retry. If nobody else has initialized the scan in the meantime, we'll
- * fill in the value we fetched on the second time through.
+ * synchronized scan machinery.
+ *
+ * If another worker initializes phs_startblock concurrently, just use
+ * their value.
*/
- if (pbscan->phs_startblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_startblock) == InvalidBlockNumber)
{
+ BlockNumber newstartblock;
+ uint32 expected = InvalidBlockNumber;
+
if (startblock != InvalidBlockNumber)
- pbscan->phs_startblock = startblock;
+ newstartblock = startblock;
else if (!pbscan->base.phs_syncscan)
- pbscan->phs_startblock = 0;
- else if (sync_startpage != InvalidBlockNumber)
- pbscan->phs_startblock = sync_startpage;
+ newstartblock = 0;
else
- {
- SpinLockRelease(&pbscan->phs_mutex);
- sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
- goto retry;
- }
+ newstartblock = ss_get_location(rel, pbscan->phs_nblocks);
+
+ pg_atomic_compare_exchange_u32(&pbscan->phs_startblock, &expected,
+ newstartblock);
}
- SpinLockRelease(&pbscan->phs_mutex);
/*
* Figure out how many blocks we're going to scan; either all of them, or
* just phs_numblock's worth, if a limit has been imposed.
*/
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* We determine the chunk size based on scan_nblocks. First we split
@@ -595,10 +590,10 @@ table_block_parallelscan_nextpage(Relation rel,
*/
/* First, figure out how many blocks we're planning on scanning */
- if (pbscan->phs_numblock == InvalidBlockNumber)
+ if (pg_atomic_read_u32(&pbscan->phs_numblock) == InvalidBlockNumber)
scan_nblocks = pbscan->phs_nblocks;
else
- scan_nblocks = pbscan->phs_numblock;
+ scan_nblocks = pg_atomic_read_u32(&pbscan->phs_numblock);
/*
* Now check if we have any remaining blocks in a previous chunk for this
@@ -644,7 +639,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (nallocated >= scan_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
- page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
+ page = (nallocated + pg_atomic_read_u32(&pbscan->phs_startblock)) % pbscan->phs_nblocks;
/*
* Report scan location. Normally, we report the current page number.
@@ -658,7 +653,7 @@ table_block_parallelscan_nextpage(Relation rel,
if (page != InvalidBlockNumber)
ss_report_location(rel, page);
else if (nallocated == pbscan->phs_nblocks)
- ss_report_location(rel, pbscan->phs_startblock);
+ ss_report_location(rel, pg_atomic_read_u32(&pbscan->phs_startblock));
}
return page;
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 2ea06a67a63..2305d0159f3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -19,7 +19,6 @@
#include "nodes/tidbitmap.h"
#include "port/atomics.h"
#include "storage/relfilelocator.h"
-#include "storage/spin.h"
#include "utils/relcache.h"
@@ -99,10 +98,9 @@ typedef struct ParallelBlockTableScanDescData
ParallelTableScanDescData base;
BlockNumber phs_nblocks; /* # blocks in relation at start of scan */
- slock_t phs_mutex; /* mutual exclusion for setting startblock */
- BlockNumber phs_startblock; /* starting block number */
- BlockNumber phs_numblock; /* # blocks to scan, or InvalidBlockNumber if
- * no limit */
+ pg_atomic_uint32 phs_startblock; /* starting block number */
+ pg_atomic_uint32 phs_numblock; /* # blocks to scan, or InvalidBlockNumber
+ * if no limit */
pg_atomic_uint64 phs_nallocated; /* number of blocks allocated to
* workers so far. */
} ParallelBlockTableScanDescData;
--
2.50.1 (Apple Git-155)
--Ot5x3ffkWEuxilWO
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment;
filename=v1-0008-convert-FastPathStrongRelationLocks-to-atomics.patch
^ permalink raw reply [nested|flat] 188+ messages in thread
end of thread, other threads:[~2026-07-09 20:21 UTC | newest]
Thread overview: 188+ messages (download: mbox mbox.gz follow: Atom feed)
-- links below jump to the message on this page --
2026-04-27 14:23 [PATCH v21 3/3] Add Assert guard to detect permission check before lock regressions Bertrand Drouvot <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[email protected]>
2026-07-09 20:21 [PATCH v1 7/8] convert ParallelBlockTableScanDescData->phs_{start,num}block to atomics Nathan Bossart <[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