agora inbox for [email protected]help / color / mirror / Atom feed
[PATCH] Allow CREATE INDEX CONCURRENTLY on partitioned table 202+ messages / 2 participants [nested] [flat]
* [PATCH] Allow CREATE INDEX CONCURRENTLY on partitioned table @ 2020-06-06 22:42 Justin Pryzby <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Justin Pryzby @ 2020-06-06 22:42 UTC (permalink / raw) Note, this effectively reverts 050098b14, so take care to not reintroduce the bug it fixed. --- doc/src/sgml/ddl.sgml | 4 +- doc/src/sgml/ref/create_index.sgml | 9 -- src/backend/commands/indexcmds.c | 214 +++++++++++++++++++------ src/include/catalog/index.h | 1 + src/test/regress/expected/indexing.out | 136 +++++++++++++++- src/test/regress/sql/indexing.sql | 26 ++- 6 files changed, 320 insertions(+), 70 deletions(-) diff --git a/doc/src/sgml/ddl.sgml b/doc/src/sgml/ddl.sgml index 03c01937094..fd56e21ef49 100644 --- a/doc/src/sgml/ddl.sgml +++ b/doc/src/sgml/ddl.sgml @@ -4131,9 +4131,7 @@ ALTER TABLE measurement ATTACH PARTITION measurement_y2008m02 so that they are applied automatically to the entire hierarchy. This is very convenient, as not only will the existing partitions become indexed, but - also any partitions that are created in the future will. One limitation is - that it's not possible to use the <literal>CONCURRENTLY</literal> - qualifier when creating such a partitioned index. To avoid long lock + also any partitions that are created in the future will. To avoid long lock times, it is possible to use <command>CREATE INDEX ON ONLY</command> the partitioned table; such an index is marked invalid, and the partitions do not get the index applied automatically. The indexes on partitions can diff --git a/doc/src/sgml/ref/create_index.sgml b/doc/src/sgml/ref/create_index.sgml index 40986aa502f..fc8cda655f0 100644 --- a/doc/src/sgml/ref/create_index.sgml +++ b/doc/src/sgml/ref/create_index.sgml @@ -692,15 +692,6 @@ Indexes: cannot. </para> - <para> - Concurrent builds for indexes on partitioned tables are currently not - supported. However, you may concurrently build the index on each - partition individually and then finally create the partitioned index - non-concurrently in order to reduce the time where writes to the - partitioned table will be locked out. In this case, building the - partitioned index is a metadata only operation. - </para> - </refsect2> </refsect1> diff --git a/src/backend/commands/indexcmds.c b/src/backend/commands/indexcmds.c index 91cee27743d..bb98e745267 100644 --- a/src/backend/commands/indexcmds.c +++ b/src/backend/commands/indexcmds.c @@ -71,6 +71,7 @@ /* non-export function prototypes */ +static void reindex_invalid_child_indexes(Oid indexRelationId); static bool CompareOpclassOptions(Datum *opts1, Datum *opts2, int natts); static void CheckPredicate(Expr *predicate); static void ComputeIndexAttrs(IndexInfo *indexInfo, @@ -104,7 +105,9 @@ static void reindex_error_callback(void *arg); static void ReindexPartitions(Oid relid, ReindexParams *params, bool isTopLevel); static void ReindexMultipleInternal(List *relids, - ReindexParams *params); + ReindexParams *params, + Oid parent, + int npart); static bool ReindexRelationConcurrently(Oid relationOid, ReindexParams *params); static void update_relispartition(Oid relationId, bool newval); @@ -697,17 +700,6 @@ DefineIndex(Oid relationId, partitioned = rel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE; if (partitioned) { - /* - * Note: we check 'stmt->concurrent' rather than 'concurrent', so that - * the error is thrown also for temporary tables. Seems better to be - * consistent, even though we could do it on temporary table because - * we're not actually doing it concurrently. - */ - if (stmt->concurrent) - ereport(ERROR, - (errcode(ERRCODE_FEATURE_NOT_SUPPORTED), - errmsg("cannot create index on partitioned table \"%s\" concurrently", - RelationGetRelationName(rel)))); if (stmt->excludeOpNames) ereport(ERROR, (errcode(ERRCODE_FEATURE_NOT_SUPPORTED), @@ -1147,6 +1139,11 @@ DefineIndex(Oid relationId, if (pd->nparts != 0) flags |= INDEX_CREATE_INVALID; } + else if (concurrent && OidIsValid(parentIndexId)) + { + /* If concurrent, initially build index partitions as "invalid" */ + flags |= INDEX_CREATE_INVALID; + } if (stmt->deferrable) constr_flags |= INDEX_CONSTR_CREATE_DEFERRABLE; @@ -1212,17 +1209,30 @@ DefineIndex(Oid relationId, partdesc = RelationGetPartitionDesc(rel, true); if ((!stmt->relation || stmt->relation->inh) && partdesc->nparts > 0) { + /* + * Need to close the relation before recursing into children, so + * copy needed data into a longlived context. + */ + + MemoryContext ind_context = AllocSetContextCreate(PortalContext, "CREATE INDEX", + ALLOCSET_DEFAULT_SIZES); + MemoryContext oldcontext = MemoryContextSwitchTo(ind_context); int nparts = partdesc->nparts; Oid *part_oids = palloc_array(Oid, nparts); bool invalidate_parent = false; Relation parentIndex; TupleDesc parentDesc; + char *relname; pgstat_progress_update_param(PROGRESS_CREATEIDX_PARTITIONS_TOTAL, nparts); /* Make a local copy of partdesc->oids[], just for safety */ memcpy(part_oids, partdesc->oids, sizeof(Oid) * nparts); + parentDesc = CreateTupleDescCopy(RelationGetDescr(rel)); + relname = pstrdup(RelationGetRelationName(rel)); + table_close(rel, NoLock); + MemoryContextSwitchTo(oldcontext); /* * We'll need an IndexInfo describing the parent index. The one @@ -1235,8 +1245,6 @@ DefineIndex(Oid relationId, parentIndex = index_open(indexRelationId, lockmode); indexInfo = BuildIndexInfo(parentIndex); - parentDesc = RelationGetDescr(rel); - /* * For each partition, scan all existing indexes; if one matches * our index definition and is not already attached to some other @@ -1276,9 +1284,9 @@ DefineIndex(Oid relationId, ereport(ERROR, (errcode(ERRCODE_WRONG_OBJECT_TYPE), errmsg("cannot create unique index on partitioned table \"%s\"", - RelationGetRelationName(rel)), + relname), errdetail("Table \"%s\" contains partitions that are foreign tables.", - RelationGetRelationName(rel)))); + relname))); AtEOXact_GUC(false, child_save_nestlevel); SetUserIdAndSecContext(child_save_userid, @@ -1364,10 +1372,14 @@ DefineIndex(Oid relationId, */ if (!found) { - IndexStmt *childStmt = copyObject(stmt); + IndexStmt *childStmt; bool found_whole_row; ListCell *lc; + oldcontext = MemoryContextSwitchTo(ind_context); + childStmt = copyObject(stmt); + MemoryContextSwitchTo(oldcontext); + /* * We can't use the same index name for the child index, * so clear idxname to let the recursive invocation choose @@ -1426,14 +1438,24 @@ DefineIndex(Oid relationId, createdConstraintId, is_alter_table, check_rights, check_not_in_use, skip_build, quiet); + SetUserIdAndSecContext(child_save_userid, child_save_sec_context); + if (concurrent) + { + PopActiveSnapshot(); + PushActiveSnapshot(GetTransactionSnapshot()); + invalidate_parent = true; + } } - pgstat_progress_update_param(PROGRESS_CREATEIDX_PARTITIONS_DONE, - i + 1); + /* For concurrent build, this is a catalog-only stage */ + if (!concurrent) + pgstat_progress_update_param(PROGRESS_CREATEIDX_PARTITIONS_DONE, + i + 1); free_attrmap(attmap); } + pfree(relname); index_close(parentIndex, lockmode); @@ -1443,46 +1465,40 @@ DefineIndex(Oid relationId, * invalid, this is incorrect, so update our row to invalid too. */ if (invalidate_parent) - { - Relation pg_index = table_open(IndexRelationId, RowExclusiveLock); - HeapTuple tup, - newtup; - - tup = SearchSysCache1(INDEXRELID, - ObjectIdGetDatum(indexRelationId)); - if (!HeapTupleIsValid(tup)) - elog(ERROR, "cache lookup failed for index %u", - indexRelationId); - newtup = heap_copytuple(tup); - ((Form_pg_index) GETSTRUCT(newtup))->indisvalid = false; - CatalogTupleUpdate(pg_index, &tup->t_self, newtup); - ReleaseSysCache(tup); - table_close(pg_index, RowExclusiveLock); - heap_freetuple(newtup); - } - } + index_set_state_flags(indexRelationId, INDEX_DROP_CLEAR_VALID); + } else + table_close(rel, NoLock); /* * Indexes on partitioned tables are not themselves built, so we're * done here. */ + AtEOXact_GUC(false, root_save_nestlevel); SetUserIdAndSecContext(root_save_userid, root_save_sec_context); - table_close(rel, NoLock); + if (!OidIsValid(parentIndexId)) + { + if (concurrent) + reindex_invalid_child_indexes(indexRelationId); + pgstat_progress_end_command(); + } + return address; } AtEOXact_GUC(false, root_save_nestlevel); SetUserIdAndSecContext(root_save_userid, root_save_sec_context); - if (!concurrent) + if (!concurrent || OidIsValid(parentIndexId)) { - /* Close the heap and we're done, in the non-concurrent case */ - table_close(rel, NoLock); + /* + * We're done if this is the top-level index, + * or the catalog-only phase of a partition built concurrently + */ - /* If this is the top-level index, we're done. */ + table_close(rel, NoLock); if (!OidIsValid(parentIndexId)) pgstat_progress_end_command(); @@ -1695,6 +1711,70 @@ DefineIndex(Oid relationId, return address; } +/* Reindex invalid child indexes created earlier */ +static void +reindex_invalid_child_indexes(Oid indexRelationId) +{ + ListCell *lc; + List *parentindexes = NIL; + int npart = 1; + ReindexParams params = { + .options = REINDEXOPT_CONCURRENTLY | REINDEXOPT_REPORT_CREATE_PART + }; + + MemoryContext ind_context = AllocSetContextCreate(PortalContext, "CREATE INDEX", + ALLOCSET_DEFAULT_SIZES); + MemoryContext oldcontext; + List *childs = find_all_inheritors(indexRelationId, ShareLock, NULL); + List *partitions = NIL; + + PreventInTransactionBlock(true, "REINDEX INDEX"); + + foreach (lc, childs) + { + Oid partoid = lfirst_oid(lc); + char partkind = get_rel_relkind(partoid); + + if (partkind == RELKIND_PARTITIONED_INDEX) + { + oldcontext = MemoryContextSwitchTo(ind_context); + parentindexes = lappend_oid(parentindexes, partoid); + MemoryContextSwitchTo(oldcontext); + } + + // npart++ ? + if (!RELKIND_HAS_STORAGE(partkind) || get_index_isvalid(partoid)) + continue; + + /* Save partition OID */ + oldcontext = MemoryContextSwitchTo(ind_context); + partitions = lappend_oid(partitions, partoid); + MemoryContextSwitchTo(oldcontext); + } + + /* + * Process each partition listed in a separate transaction. Note that + * this commits and then starts a new transaction immediately. + * XXX: since this is done in 2*N transactions, it could just as well + * call ReindexRelationConcurrently directly + */ + ReindexMultipleInternal(partitions, ¶ms, indexRelationId, npart); + + /* + * CIC needs to mark a partitioned index as VALID, which itself + * requires setting READY, which is unset for CIC (even though it's + * meaningless for an index without storage). This must be done only + * while holding a lock which precludes adding partitions. + * See also: validatePartitionedIndex(). + */ + foreach (lc, parentindexes) + { + Oid partoid = lfirst_oid(lc); + index_set_state_flags(partoid, INDEX_CREATE_SET_READY); + CommandCounterIncrement(); + index_set_state_flags(partoid, INDEX_CREATE_SET_VALID); + } +} /* * CheckMutability @@ -3084,7 +3164,7 @@ ReindexMultipleTables(const char *objectName, ReindexObjectType objectKind, * Process each relation listed in a separate transaction. Note that this * commits and then starts a new transaction immediately. */ - ReindexMultipleInternal(relids, params); + ReindexMultipleInternal(relids, params, InvalidOid, 0); MemoryContextDelete(private_context); } @@ -3120,6 +3200,7 @@ ReindexPartitions(Oid relid, ReindexParams *params, bool isTopLevel) char relkind = get_rel_relkind(relid); char *relname = get_rel_name(relid); char *relnamespace = get_namespace_name(get_rel_namespace(relid)); + int npart = 1; MemoryContext reindex_context; List *inhoids; ListCell *lc; @@ -3190,7 +3271,7 @@ ReindexPartitions(Oid relid, ReindexParams *params, bool isTopLevel) * Process each partition listed in a separate transaction. Note that * this commits and then starts a new transaction immediately. */ - ReindexMultipleInternal(partitions, params); + ReindexMultipleInternal(partitions, params, relid, npart); /* * Clean up working storage --- note we must do this after @@ -3208,7 +3289,7 @@ ReindexPartitions(Oid relid, ReindexParams *params, bool isTopLevel) * and starts a new transaction when finished. */ static void -ReindexMultipleInternal(List *relids, ReindexParams *params) +ReindexMultipleInternal(List *relids, ReindexParams *params, Oid parent, int npart) { ListCell *l; @@ -3302,6 +3383,29 @@ ReindexMultipleInternal(List *relids, ReindexParams *params) } CommitTransactionCommand(); + + if (params->options & REINDEXOPT_REPORT_CREATE_PART) + { + /* + * Update pgstat progress report to indicate that create index on + * partition was finished. + */ + const int progress_cols[] = { + PROGRESS_CREATEIDX_COMMAND, + PROGRESS_CREATEIDX_INDEX_OID, + PROGRESS_CREATEIDX_PHASE, + PROGRESS_CREATEIDX_PARTITIONS_DONE + }; + const int64 progress_vals[] = { + PROGRESS_CREATEIDX_COMMAND_CREATE_CONCURRENTLY, + parent, + PROGRESS_CREATEIDX_PHASE_VALIDATE_IDXSCAN, + npart++ + }; + + pgstat_progress_update_multi_param(4, progress_cols, progress_vals); + } + } StartTransactionCommand(); @@ -3697,7 +3801,9 @@ ReindexRelationConcurrently(Oid relationOid, ReindexParams *params) if (indexRel->rd_rel->relpersistence == RELPERSISTENCE_TEMP) elog(ERROR, "cannot reindex a temporary table concurrently"); - pgstat_progress_start_command(PROGRESS_COMMAND_CREATE_INDEX, + /* Don't overwrite CREATE INDEX command */ + if (!(params->options & REINDEXOPT_REPORT_CREATE_PART)) + pgstat_progress_start_command(PROGRESS_COMMAND_CREATE_INDEX, idx->tableId); progress_vals[0] = PROGRESS_CREATEIDX_COMMAND_REINDEX_CONCURRENTLY; @@ -3857,9 +3963,11 @@ ReindexRelationConcurrently(Oid relationOid, ReindexParams *params) /* * Update progress for the index to build, with the correct parent - * table involved. + * table involved. Don't overwrite CREATE INDEX command. */ - pgstat_progress_start_command(PROGRESS_COMMAND_CREATE_INDEX, newidx->tableId); + if (!(params->options & REINDEXOPT_REPORT_CREATE_PART)) + pgstat_progress_start_command(PROGRESS_COMMAND_CREATE_INDEX, newidx->tableId); + progress_vals[0] = PROGRESS_CREATEIDX_COMMAND_REINDEX_CONCURRENTLY; progress_vals[1] = PROGRESS_CREATEIDX_PHASE_BUILD; progress_vals[2] = newidx->indexId; @@ -3921,10 +4029,12 @@ ReindexRelationConcurrently(Oid relationOid, ReindexParams *params) /* * Update progress for the index to build, with the correct parent - * table involved. + * table involved. Don't overwrite CREATE INDEX command. */ - pgstat_progress_start_command(PROGRESS_COMMAND_CREATE_INDEX, + if (!(params->options & REINDEXOPT_REPORT_CREATE_PART)) + pgstat_progress_start_command(PROGRESS_COMMAND_CREATE_INDEX, newidx->tableId); + progress_vals[0] = PROGRESS_CREATEIDX_COMMAND_REINDEX_CONCURRENTLY; progress_vals[1] = PROGRESS_CREATEIDX_PHASE_VALIDATE_IDXSCAN; progress_vals[2] = newidx->indexId; @@ -4159,7 +4269,9 @@ ReindexRelationConcurrently(Oid relationOid, ReindexParams *params) MemoryContextDelete(private_context); - pgstat_progress_end_command(); + /* Don't overwrite CREATE INDEX command. */ + if (!(params->options & REINDEXOPT_REPORT_CREATE_PART)) + pgstat_progress_end_command(); return true; } diff --git a/src/include/catalog/index.h b/src/include/catalog/index.h index 91c28868d47..4142894c230 100644 --- a/src/include/catalog/index.h +++ b/src/include/catalog/index.h @@ -42,6 +42,7 @@ typedef struct ReindexParams #define REINDEXOPT_REPORT_PROGRESS 0x02 /* report pgstat progress */ #define REINDEXOPT_MISSING_OK 0x04 /* skip missing relations */ #define REINDEXOPT_CONCURRENTLY 0x08 /* concurrent mode */ +#define REINDEXOPT_REPORT_CREATE_PART 0x10 /* report that index was created for partition */ /* state info for validate_index bulkdelete callback */ typedef struct ValidateIndexState diff --git a/src/test/regress/expected/indexing.out b/src/test/regress/expected/indexing.out index 1bdd430f063..6b2320244b8 100644 --- a/src/test/regress/expected/indexing.out +++ b/src/test/regress/expected/indexing.out @@ -50,11 +50,139 @@ select relname, relkind, relhassubclass, inhparent::regclass (8 rows) drop table idxpart; --- Some unsupported features +-- CIC on partitioned table create table idxpart (a int, b int, c text) partition by range (a); -create table idxpart1 partition of idxpart for values from (0) to (10); -create index concurrently on idxpart (a); -ERROR: cannot create index on partitioned table "idxpart" concurrently +create table idxpart1 partition of idxpart for values from (0) to (10) partition by range(a); +create table idxpart11 partition of idxpart1 for values from (0) to (10) partition by range(a); +create table idxpart111 partition of idxpart11 default partition by range(a); +create table idxpart1111 partition of idxpart111 default partition by range(a); +create table idxpart2 partition of idxpart for values from (10) to (20); +create table idxpart3 partition of idxpart for values from (30) to (40) partition by range(a); +create table idxpart31 partition of idxpart3 default; +insert into idxpart2 values(10),(10); -- not unique +create index concurrently on idxpart (a); -- partitioned +create index concurrently on idxpart1 (a); -- partitioned and partition +create index concurrently on idxpart11 (a); -- partitioned and partition, with no leaves +create index concurrently on idxpart2 (a); -- leaf +create unique index concurrently on idxpart (a); -- partitioned, unique failure +ERROR: could not create unique index "idxpart2_a_idx2_ccnew" +DETAIL: Key (a)=(10) is duplicated. +\d idxpart + Partitioned table "public.idxpart" + Column | Type | Collation | Nullable | Default +--------+---------+-----------+----------+--------- + a | integer | | | + b | integer | | | + c | text | | | +Partition key: RANGE (a) +Indexes: + "idxpart_a_idx" btree (a) + "idxpart_a_idx1" UNIQUE, btree (a) INVALID +Number of partitions: 3 (Use \d+ to list them.) + +\d idxpart1 + Partitioned table "public.idxpart1" + Column | Type | Collation | Nullable | Default +--------+---------+-----------+----------+--------- + a | integer | | | + b | integer | | | + c | text | | | +Partition of: idxpart FOR VALUES FROM (0) TO (10) +Partition key: RANGE (a) +Indexes: + "idxpart1_a_idx" btree (a) + "idxpart1_a_idx1" btree (a) + "idxpart1_a_idx2" UNIQUE, btree (a) INVALID +Number of partitions: 1 (Use \d+ to list them.) + +\d idxpart11 + Partitioned table "public.idxpart11" + Column | Type | Collation | Nullable | Default +--------+---------+-----------+----------+--------- + a | integer | | | + b | integer | | | + c | text | | | +Partition of: idxpart1 FOR VALUES FROM (0) TO (10) +Partition key: RANGE (a) +Indexes: + "idxpart11_a_idx" btree (a) + "idxpart11_a_idx1" btree (a) + "idxpart11_a_idx2" btree (a) + "idxpart11_a_idx3" UNIQUE, btree (a) INVALID +Number of partitions: 1 (Use \d+ to list them.) + +\d idxpart111 + Partitioned table "public.idxpart111" + Column | Type | Collation | Nullable | Default +--------+---------+-----------+----------+--------- + a | integer | | | + b | integer | | | + c | text | | | +Partition of: idxpart11 DEFAULT +Partition key: RANGE (a) +Indexes: + "idxpart111_a_idx" btree (a) + "idxpart111_a_idx1" btree (a) + "idxpart111_a_idx2" btree (a) + "idxpart111_a_idx3" UNIQUE, btree (a) INVALID +Number of partitions: 1 (Use \d+ to list them.) + +\d idxpart1111 + Partitioned table "public.idxpart1111" + Column | Type | Collation | Nullable | Default +--------+---------+-----------+----------+--------- + a | integer | | | + b | integer | | | + c | text | | | +Partition of: idxpart111 DEFAULT +Partition key: RANGE (a) +Indexes: + "idxpart1111_a_idx" btree (a) + "idxpart1111_a_idx1" btree (a) + "idxpart1111_a_idx2" btree (a) + "idxpart1111_a_idx3" UNIQUE, btree (a) INVALID +Number of partitions: 0 + +\d idxpart2 + Table "public.idxpart2" + Column | Type | Collation | Nullable | Default +--------+---------+-----------+----------+--------- + a | integer | | | + b | integer | | | + c | text | | | +Partition of: idxpart FOR VALUES FROM (10) TO (20) +Indexes: + "idxpart2_a_idx" btree (a) + "idxpart2_a_idx1" btree (a) + "idxpart2_a_idx2" UNIQUE, btree (a) INVALID + "idxpart2_a_idx2_ccnew" UNIQUE, btree (a) INVALID + +\d idxpart3 + Partitioned table "public.idxpart3" + Column | Type | Collation | Nullable | Default +--------+---------+-----------+----------+--------- + a | integer | | | + b | integer | | | + c | text | | | +Partition of: idxpart FOR VALUES FROM (30) TO (40) +Partition key: RANGE (a) +Indexes: + "idxpart3_a_idx" btree (a) + "idxpart3_a_idx1" UNIQUE, btree (a) INVALID +Number of partitions: 1 (Use \d+ to list them.) + +\d idxpart31 + Table "public.idxpart31" + Column | Type | Collation | Nullable | Default +--------+---------+-----------+----------+--------- + a | integer | | | + b | integer | | | + c | text | | | +Partition of: idxpart3 DEFAULT +Indexes: + "idxpart31_a_idx" btree (a) + "idxpart31_a_idx1" UNIQUE, btree (a) INVALID + drop table idxpart; -- Verify bugfix with query on indexed partitioned table with no partitions -- https://postgr.es/m/[email protected] diff --git a/src/test/regress/sql/indexing.sql b/src/test/regress/sql/indexing.sql index 429120e7104..14e0513386c 100644 --- a/src/test/regress/sql/indexing.sql +++ b/src/test/regress/sql/indexing.sql @@ -29,10 +29,30 @@ select relname, relkind, relhassubclass, inhparent::regclass where relname like 'idxpart%' order by relname; drop table idxpart; --- Some unsupported features +-- CIC on partitioned table create table idxpart (a int, b int, c text) partition by range (a); -create table idxpart1 partition of idxpart for values from (0) to (10); -create index concurrently on idxpart (a); +create table idxpart1 partition of idxpart for values from (0) to (10) partition by range(a); +create table idxpart11 partition of idxpart1 for values from (0) to (10) partition by range(a); +create table idxpart111 partition of idxpart11 default partition by range(a); +create table idxpart1111 partition of idxpart111 default partition by range(a); +create table idxpart2 partition of idxpart for values from (10) to (20); +create table idxpart3 partition of idxpart for values from (30) to (40) partition by range(a); +create table idxpart31 partition of idxpart3 default; + +insert into idxpart2 values(10),(10); -- not unique +create index concurrently on idxpart (a); -- partitioned +create index concurrently on idxpart1 (a); -- partitioned and partition +create index concurrently on idxpart11 (a); -- partitioned and partition, with no leaves +create index concurrently on idxpart2 (a); -- leaf +create unique index concurrently on idxpart (a); -- partitioned, unique failure +\d idxpart +\d idxpart1 +\d idxpart11 +\d idxpart111 +\d idxpart1111 +\d idxpart2 +\d idxpart3 +\d idxpart31 drop table idxpart; -- Verify bugfix with query on indexed partitioned table with no partitions -- 2.25.1 --aPdhxNJGSeOG9wFI-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v2 3/3] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 150 ++++++++++++++++++++++++++-- 2 files changed, 142 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fe470de63f2..7b245daaf4e 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,46 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +/* + * Check if the buffer is already undergoing read AIO. If it is, assign the + * IO's wait reference to operation->io_wref, thereby allowing the caller to + * wait for that IO. + */ +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if ((buf_state & BM_IO_IN_PROGRESS) && !(buf_state & BM_VALID)) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if ((buf_state & BM_IO_IN_PROGRESS) && !(buf_state & BM_VALID)) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1729,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1748,66 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (unlikely(operation->foreign_io)) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + + /* + * Report and track this as a 'hit' for this backend, even + * though it must have started out as a miss in + * PinBufferForBlock(). The other backend (or ourselves, + * as part of a read started earlier) will track this as a + * 'read'. + */ + TRACE_POSTGRESQL_BUFFER_READ_DONE(operation->forknum, + operation->blocknum + operation->nblocks_done, + operation->smgr->smgr_rlocator.locator.spcOid, + operation->smgr->smgr_rlocator.locator.dbOid, + operation->smgr->smgr_rlocator.locator.relNumber, + operation->smgr->smgr_rlocator.backend, + true); + + if (BufferIsLocal(buffer)) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1893,43 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses of + * an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + * + * It's possible that another backend starts IO on the buffer between this + * check and the ReadBuffersCanStartIO(nowait = false) below. In that case + * we will synchronously wait for the IO below, but the window for that is + * small enough that it won't happen often enough to have a significant + * performance impact. + */ + if (ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + *nblocks_progress = 1; + operation->foreign_io = true; + + CheckReadBuffersOperation(operation, false); + + + return true; + } + + operation->foreign_io = false; + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1987,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O is in progress in another backend (it can't be + * this backend), we want to wait for the outcome: either done, or + * something went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { -- 2.48.1.76.g4e746b1a31.dirty --aknbaqiolq3mrqjk-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <[email protected]> 0 siblings, 0 replies; 202+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 202+ messages in thread
end of thread, other threads:[~2025-08-15 15:01 UTC | newest] Thread overview: 202+ messages (download: mbox mbox.gz follow: Atom feed) -- links below jump to the message on this page -- 2020-06-06 22:42 [PATCH] Allow CREATE INDEX CONCURRENTLY on partitioned table Justin Pryzby <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v2 3/3] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
This inbox is served by agora; see mirroring instructions for how to clone and mirror all data and code used for this inbox