agora inbox for pgsql-hackers@postgresql.orghelp / color / mirror / Atom feed
[PATCH v2 2/7] Row pattern recognition patch (parse/analysis). 332+ messages / 2 participants [nested] [flat]
* [PATCH v2 2/7] Row pattern recognition patch (parse/analysis). @ 2023-06-26 08:05 Tatsuo Ishii <ishii@postgresql.org> 0 siblings, 0 replies; 332+ messages in thread From: Tatsuo Ishii @ 2023-06-26 08:05 UTC (permalink / raw) --- src/backend/parser/parse_agg.c | 7 ++ src/backend/parser/parse_clause.c | 192 +++++++++++++++++++++++++++++- src/backend/parser/parse_expr.c | 4 + src/backend/parser/parse_func.c | 3 + 4 files changed, 205 insertions(+), 1 deletion(-) diff --git a/src/backend/parser/parse_agg.c b/src/backend/parser/parse_agg.c index 85cd47b7ae..aa7a1cee80 100644 --- a/src/backend/parser/parse_agg.c +++ b/src/backend/parser/parse_agg.c @@ -564,6 +564,10 @@ check_agglevels_and_constraints(ParseState *pstate, Node *expr) errkind = true; break; + case EXPR_KIND_RPR_DEFINE: + errkind = true; + break; + /* * There is intentionally no default: case here, so that the * compiler will warn if we add a new ParseExprKind without @@ -953,6 +957,9 @@ transformWindowFuncCall(ParseState *pstate, WindowFunc *wfunc, case EXPR_KIND_CYCLE_MARK: errkind = true; break; + case EXPR_KIND_RPR_DEFINE: + errkind = true; + break; /* * There is intentionally no default: case here, so that the diff --git a/src/backend/parser/parse_clause.c b/src/backend/parser/parse_clause.c index f61f794755..ccf3332bef 100644 --- a/src/backend/parser/parse_clause.c +++ b/src/backend/parser/parse_clause.c @@ -100,7 +100,10 @@ static WindowClause *findWindowClause(List *wclist, const char *name); static Node *transformFrameOffset(ParseState *pstate, int frameOptions, Oid rangeopfamily, Oid rangeopcintype, Oid *inRangeFunc, Node *clause); - +static void transformRPR(ParseState *pstate, WindowClause *wc, WindowDef *windef); +static List *transformDefineClause(ParseState *pstate, WindowClause *wc, WindowDef *windef); +static List *transformPatternClause(ParseState *pstate, WindowClause *wc, WindowDef *windef); +static List *transformMeasureClause(ParseState *pstate, WindowClause *wc, WindowDef *windef); /* * transformFromClause - @@ -2949,6 +2952,10 @@ transformWindowDefinitions(ParseState *pstate, rangeopfamily, rangeopcintype, &wc->endInRangeFunc, windef->endOffset); + + /* Process Row Pattern Recognition related clauses */ + transformRPR(pstate, wc, windef); + wc->runCondition = NIL; wc->winref = winref; @@ -3814,3 +3821,186 @@ transformFrameOffset(ParseState *pstate, int frameOptions, return node; } + +/* + * transformRPR + * Process Row Pattern Recognition related clauses + */ +static void +transformRPR(ParseState *pstate, WindowClause *wc, WindowDef *windef) +{ + /* Check Frame option. Frame must start at current row */ + + /* + * Window definition exists? + */ + if (windef == NULL) + return; + + /* + * Row Pattern Common Syntax clause exists? + */ + if (windef->rpCommonSyntax == NULL) + return; + + if ((wc->frameOptions & FRAMEOPTION_START_CURRENT_ROW) == 0) + ereport(ERROR, + (errcode(ERRCODE_SYNTAX_ERROR), + errmsg("FRAME must start at current row when row patttern recognition is used"))); + + /* Transform AFTER MACH SKIP TO clause */ + wc->rpSkipTo = windef->rpCommonSyntax->rpSkipTo; + + /* Transform SEEK or INITIAL clause */ + wc->initial = windef->rpCommonSyntax->initial; + + /* Transform DEFINE clause into list of TargetEntry's */ + wc->defineClause = transformDefineClause(pstate, wc, windef); + + /* Check PATTERN clause and copy to patternClause */ + transformPatternClause(pstate, wc, windef); + + /* Transform MEASURE clause */ + transformMeasureClause(pstate, wc, windef); +} + +/* + * transformDefineClause Process DEFINE clause and transform ResTarget into + * list of TargetEntry. + * + * XXX we only support column reference in row pattern definition search + * condition, e.g. "price". <row pattern definition variable name>.<column + * reference> is not supported, e.g. "A.price". + */ +static List * +transformDefineClause(ParseState *pstate, WindowClause *wc, WindowDef *windef) +{ + ListCell *lc; + ResTarget *restarget, *r; + List *restargets; + + + /* + * If Row Definition Common Syntax exists, DEFINE clause must exist. + * (the raw parser should have already checked it.) + */ + Assert(windef->rpCommonSyntax->rpDefs != NULL); + + /* + * Check for duplicate row pattern definition variables. The standard + * requires that no two row pattern definition variable names shall be + * equivalent. + */ + restargets = NIL; + foreach(lc, windef->rpCommonSyntax->rpDefs) + { + char *name; + ListCell *l; + + restarget = (ResTarget *)lfirst(lc); + name = restarget->name; + + /* + * Make sure that row pattern definition search condition is a boolean + * expression. + */ + transformWhereClause(pstate, restarget->val, + EXPR_KIND_RPR_DEFINE, "DEFINE"); + + foreach(l, restargets) + { + char *n; + + r = (ResTarget *) lfirst(l); + n = r->name; + + if (!strcmp(n, name)) + ereport(ERROR, + (errcode(ERRCODE_SYNTAX_ERROR), + errmsg("row pattern definition variable name \"%s\" appears more than once in DEFINE clause", + name), + parser_errposition(pstate, exprLocation((Node *)r)))); + } + restargets = lappend(restargets, restarget); + } + list_free(restargets); + + return transformTargetList(pstate, windef->rpCommonSyntax->rpDefs, + EXPR_KIND_RPR_DEFINE); +} + +/* + * transformPatternClause + * Process PATTERN clause and return PATTERN clause in the raw parse tree + */ +static List * +transformPatternClause(ParseState *pstate, WindowClause *wc, WindowDef *windef) +{ + List *patterns; + ListCell *lc, *l; + + /* + * Row Pattern Common Syntax clause exists? + */ + if (windef->rpCommonSyntax == NULL) + return NULL; + + /* + * Primary row pattern variable names in PATTERN clause must appear in + * DEFINE clause as row pattern definition variable names. + */ + wc->patternVariable = NIL; + wc->patternRegexp = NIL; + foreach(lc, windef->rpCommonSyntax->rpPatterns) + { + A_Expr *a; + char *name; + char *regexp; + bool found = false; + + if (!IsA(lfirst(lc), A_Expr)) + ereport(ERROR, + errmsg("node type is not A_Expr")); + + a = (A_Expr *)lfirst(lc); + name = strVal(a->lexpr); + + foreach(l, windef->rpCommonSyntax->rpDefs) + { + ResTarget *restarget = (ResTarget *)lfirst(l); + + if (!strcmp(restarget->name, name)) + { + found = true; + break; + } + } + + if (!found) + ereport(ERROR, + (errcode(ERRCODE_SYNTAX_ERROR), + errmsg("primary row pattern variable name \"%s\" does not appear in DEFINE clause", + name), + parser_errposition(pstate, exprLocation((Node *)a)))); + wc->patternVariable = lappend(wc->patternVariable, makeString(pstrdup(name))); + regexp = strVal(lfirst(list_head(a->name))); + wc->patternRegexp = lappend(wc->patternRegexp, makeString(pstrdup(regexp))); + } + return patterns; +} + +/* + * transformMeasureClause + * Process MEASURE clause + */ +static List * +transformMeasureClause(ParseState *pstate, WindowClause *wc, WindowDef *windef) +{ + if (windef->rowPatternMeasures == NIL) + return NIL; + + ereport(ERROR, + (errcode(ERRCODE_SYNTAX_ERROR), + errmsg("%s","MEASURE clause is not supported yet"), + parser_errposition(pstate, exprLocation((Node *)windef->rowPatternMeasures)))); +} diff --git a/src/backend/parser/parse_expr.c b/src/backend/parser/parse_expr.c index 346fd272b6..20231d9ec0 100644 --- a/src/backend/parser/parse_expr.c +++ b/src/backend/parser/parse_expr.c @@ -541,6 +541,7 @@ transformColumnRef(ParseState *pstate, ColumnRef *cref) case EXPR_KIND_COPY_WHERE: case EXPR_KIND_GENERATED_COLUMN: case EXPR_KIND_CYCLE_MARK: + case EXPR_KIND_RPR_DEFINE: /* okay */ break; @@ -1754,6 +1755,7 @@ transformSubLink(ParseState *pstate, SubLink *sublink) case EXPR_KIND_VALUES: case EXPR_KIND_VALUES_SINGLE: case EXPR_KIND_CYCLE_MARK: + case EXPR_KIND_RPR_DEFINE: /* okay */ break; case EXPR_KIND_CHECK_CONSTRAINT: @@ -3133,6 +3135,8 @@ ParseExprKindName(ParseExprKind exprKind) return "GENERATED AS"; case EXPR_KIND_CYCLE_MARK: return "CYCLE"; + case EXPR_KIND_RPR_DEFINE: + return "DEFINE"; /* * There is intentionally no default: case here, so that the diff --git a/src/backend/parser/parse_func.c b/src/backend/parser/parse_func.c index b3f0b6a137..2ff3699538 100644 --- a/src/backend/parser/parse_func.c +++ b/src/backend/parser/parse_func.c @@ -2656,6 +2656,9 @@ check_srf_call_placement(ParseState *pstate, Node *last_srf, int location) case EXPR_KIND_CYCLE_MARK: errkind = true; break; + case EXPR_KIND_RPR_DEFINE: + errkind = true; + break; /* * There is intentionally no default: case here, so that the -- 2.25.1 ----Next_Part(Mon_Jun_26_17_45_07_2023_724)-- Content-Type: Text/X-Patch; charset=us-ascii Content-Transfer-Encoding: 7bit Content-Disposition: inline; filename="v2-0003-Row-pattern-recognition-patch-planner.patch" ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ 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 <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 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] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO @ 2025-08-15 15:01 Andres Freund <andres@anarazel.de> 0 siblings, 0 replies; 332+ messages in thread From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw) Author: Reviewed-by: Discussion: https://postgr.es/m/ Backpatch: --- src/include/storage/bufmgr.h | 1 + src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 9 deletions(-) diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h index 41fdc1e7693..7ddb867bc99 100644 --- a/src/include/storage/bufmgr.h +++ b/src/include/storage/bufmgr.h @@ -137,6 +137,7 @@ struct ReadBuffersOperation int flags; int16 nblocks; int16 nblocks_done; + bool foreign_io; PgAioWaitRef io_wref; PgAioReturn io_return; }; diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c index fd7e21d96d3..de755fd53ad 100644 --- a/src/backend/storage/buffer/bufmgr.c +++ b/src/backend/storage/buffer/bufmgr.c @@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait) return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait); } +static inline bool +ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer) +{ + BufferDesc *desc; + uint32 buf_state; + PgAioWaitRef iow; + + pgaio_wref_clear(&iow); + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + + if (buf_state & BM_IO_IN_PROGRESS) + iow = desc->io_wref; + UnlockBufHdr(desc, buf_state); + } + + if (pgaio_wref_valid(&iow)) + { + operation->io_wref = iow; + return true; + } + + return false; +} + /* * Helper for AsyncReadBuffers that tries to get the buffer ready for IO. */ @@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation) * * we first check if we already know the IO is complete. */ - if (aio_ret->result.status == PGAIO_RS_UNKNOWN && + if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) && !pgaio_wref_check_done(&operation->io_wref)) { instr_time io_start = pgstat_prepare_io_time(track_io_timing); @@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation) Assert(pgaio_wref_check_done(&operation->io_wref)); } - /* - * We now are sure the IO completed. Check the results. This - * includes reporting on errors if there were any. - */ - ProcessReadBuffersResult(operation); + if (operation->foreign_io) + { + Buffer buffer = operation->buffers[operation->nblocks_done]; + BufferDesc *desc; + uint32 buf_state; + + if (BufferIsLocal(buffer)) + { + desc = GetLocalBufferDescriptor(-buffer - 1); + buf_state = pg_atomic_read_u32(&desc->state); + } + else + { + desc = GetBufferDescriptor(buffer - 1); + buf_state = LockBufHdr(desc); + UnlockBufHdr(desc, buf_state); + } + + if (buf_state & BM_VALID) + { + operation->nblocks_done += 1; + Assert(operation->nblocks_done <= operation->nblocks); + } + } + else + { + /* + * We now are sure the IO completed. Check the results. This + * includes reporting on errors if there were any. + */ + ProcessReadBuffersResult(operation); + } } /* @@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) io_object = IOOBJECT_RELATION; } + /* + * If AIO is in progress, be it in this backend or another backend, we + * just associate the wait reference with the operation and wait in + * WaitReadBuffers(). This turns out to be important for performance in + * two workloads: + * + * 1) A read stream that has to read the same block multiple times within + * the readahead distance. This can happen e.g. for the table accesses + * of an index scan. + * + * 2) Concurrent scans by multiple backends on the same relation. + * + * If we were to synchronously wait for the in-progress IO, we'd not be + * able to keep enough I/O in flight. + * + * + * If we do find there is ongoing I/O for the buffer, we set up a 1-block + * ReadBuffersOperation that WaitReadBuffers then can wait on. + */ + if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done])) + { + /* FIXME: probably need to wait if io_method == sync? */ + + *nblocks_progress = 1; + did_start_io = true; + operation->foreign_io = true; + + if (0) + elog(LOG, "using foreign IO path"); + + /* FIXME: trace point */ + + /* + * FIXME: how should this be accounted for in stats? Account as a hit + * for now, quite likely *we* started this IO. + */ + if (persistence == RELPERSISTENCE_TEMP) + pgBufferUsage.local_blks_hit += 1; + else + pgBufferUsage.shared_blks_hit += 1; + + if (operation->rel) + pgstat_count_buffer_hit(operation->rel); + pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0); + if (VacuumCostActive) + VacuumCostBalance += VacuumCostPageHit; + + return true; + } + /* * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR * flag. The reason for that is that, hopefully, zero_damaged_pages isn't @@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) /* * Check if we can start IO on the first to-be-read buffer. * - * If an I/O is already in progress in another backend, we want to wait - * for the outcome: either done, or something went wrong and we will - * retry. + * If a synchronous I/O in progress in another backend (it can't be this + * backend), we want to wait for the outcome: either done, or something + * went wrong and we will retry. */ if (!ReadBuffersCanStartIO(buffers[nblocks_done], false)) { @@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress) if (VacuumCostActive) VacuumCostBalance += VacuumCostPageMiss * io_buffers_len; + operation->foreign_io = false; *nblocks_progress = io_buffers_len; did_start_io = true; } @@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf) */ if (pgaio_wref_valid(&iow)) { + if (0) + elog(LOG, "foreign wait"); pgaio_wref_wait(&iow); /* -- 2.48.1.76.g4e746b1a31.dirty --j4imdtjxqt5tb6mo-- ^ permalink raw reply [nested|flat] 332+ messages in thread
end of thread, other threads:[~2025-08-15 15:01 UTC | newest] Thread overview: 332+ messages (download: mbox mbox.gz follow: Atom feed) -- links below jump to the message on this page -- 2023-06-26 08:05 [PATCH v2 2/7] Row pattern recognition patch (parse/analysis). Tatsuo Ishii <ishii@postgresql.org> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v2 3/3] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de> 2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <andres@anarazel.de>
This inbox is served by agora; see mirroring instructions for how to clone and mirror all data and code used for this inbox