Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * execParallel.c
4 : * Support routines for parallel execution.
5 : *
6 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : * This file contains routines that are intended to support setting up,
10 : * using, and tearing down a ParallelContext from within the PostgreSQL
11 : * executor. The ParallelContext machinery will handle starting the
12 : * workers and ensuring that their state generally matches that of the
13 : * leader; see src/backend/access/transam/README.parallel for details.
14 : * However, we must save and restore relevant executor state, such as
15 : * any ParamListInfo associated with the query, buffer/WAL usage info, and
16 : * the actual plan to be passed down to the worker.
17 : *
18 : * IDENTIFICATION
19 : * src/backend/executor/execParallel.c
20 : *
21 : *-------------------------------------------------------------------------
22 : */
23 :
24 : #include "postgres.h"
25 :
26 : #include "executor/execParallel.h"
27 : #include "executor/executor.h"
28 : #include "executor/nodeAgg.h"
29 : #include "executor/nodeAppend.h"
30 : #include "executor/nodeBitmapHeapscan.h"
31 : #include "executor/nodeBitmapIndexscan.h"
32 : #include "executor/nodeCustom.h"
33 : #include "executor/nodeForeignscan.h"
34 : #include "executor/nodeHash.h"
35 : #include "executor/nodeHashjoin.h"
36 : #include "executor/nodeIncrementalSort.h"
37 : #include "executor/nodeIndexonlyscan.h"
38 : #include "executor/nodeIndexscan.h"
39 : #include "executor/nodeMemoize.h"
40 : #include "executor/nodeSeqscan.h"
41 : #include "executor/nodeSort.h"
42 : #include "executor/nodeSubplan.h"
43 : #include "executor/nodeTidrangescan.h"
44 : #include "executor/tqueue.h"
45 : #include "jit/jit.h"
46 : #include "nodes/nodeFuncs.h"
47 : #include "pgstat.h"
48 : #include "storage/proc.h"
49 : #include "tcop/tcopprot.h"
50 : #include "utils/datum.h"
51 : #include "utils/dsa.h"
52 : #include "utils/lsyscache.h"
53 : #include "utils/snapmgr.h"
54 :
55 : /*
56 : * Magic numbers for parallel executor communication. We use constants
57 : * greater than any 32-bit integer here so that values < 2^32 can be used
58 : * by individual parallel nodes to store their own state.
59 : */
60 : #define PARALLEL_KEY_EXECUTOR_FIXED UINT64CONST(0xE000000000000001)
61 : #define PARALLEL_KEY_PLANNEDSTMT UINT64CONST(0xE000000000000002)
62 : #define PARALLEL_KEY_PARAMLISTINFO UINT64CONST(0xE000000000000003)
63 : #define PARALLEL_KEY_BUFFER_USAGE UINT64CONST(0xE000000000000004)
64 : #define PARALLEL_KEY_TUPLE_QUEUE UINT64CONST(0xE000000000000005)
65 : #define PARALLEL_KEY_INSTRUMENTATION UINT64CONST(0xE000000000000006)
66 : #define PARALLEL_KEY_DSA UINT64CONST(0xE000000000000007)
67 : #define PARALLEL_KEY_QUERY_TEXT UINT64CONST(0xE000000000000008)
68 : #define PARALLEL_KEY_JIT_INSTRUMENTATION UINT64CONST(0xE000000000000009)
69 : #define PARALLEL_KEY_WAL_USAGE UINT64CONST(0xE00000000000000A)
70 :
71 : #define PARALLEL_TUPLE_QUEUE_SIZE 65536
72 :
73 : /*
74 : * Fixed-size random stuff that we need to pass to parallel workers.
75 : */
76 : typedef struct FixedParallelExecutorState
77 : {
78 : int64 tuples_needed; /* tuple bound, see ExecSetTupleBound */
79 : dsa_pointer param_exec;
80 : int eflags;
81 : int jit_flags;
82 : } FixedParallelExecutorState;
83 :
84 : /*
85 : * DSM structure for accumulating per-PlanState instrumentation.
86 : *
87 : * instrument_options: Same meaning here as in instrument.c.
88 : *
89 : * instrument_offset: Offset, relative to the start of this structure,
90 : * of the first Instrumentation object. This will depend on the length of
91 : * the plan_node_id array.
92 : *
93 : * num_workers: Number of workers.
94 : *
95 : * num_plan_nodes: Number of plan nodes.
96 : *
97 : * plan_node_id: Array of plan nodes for which we are gathering instrumentation
98 : * from parallel workers. The length of this array is given by num_plan_nodes.
99 : */
100 : struct SharedExecutorInstrumentation
101 : {
102 : int instrument_options;
103 : int instrument_offset;
104 : int num_workers;
105 : int num_plan_nodes;
106 : int plan_node_id[FLEXIBLE_ARRAY_MEMBER];
107 : /* array of num_plan_nodes * num_workers Instrumentation objects follows */
108 : };
109 : #define GetInstrumentationArray(sei) \
110 : (StaticAssertVariableIsOfTypeMacro(sei, SharedExecutorInstrumentation *), \
111 : (Instrumentation *) (((char *) sei) + sei->instrument_offset))
112 :
113 : /* Context object for ExecParallelEstimate. */
114 : typedef struct ExecParallelEstimateContext
115 : {
116 : ParallelContext *pcxt;
117 : int nnodes;
118 : } ExecParallelEstimateContext;
119 :
120 : /* Context object for ExecParallelInitializeDSM. */
121 : typedef struct ExecParallelInitializeDSMContext
122 : {
123 : ParallelContext *pcxt;
124 : SharedExecutorInstrumentation *instrumentation;
125 : int nnodes;
126 : } ExecParallelInitializeDSMContext;
127 :
128 : /* Helper functions that run in the parallel leader. */
129 : static char *ExecSerializePlan(Plan *plan, EState *estate);
130 : static bool ExecParallelEstimate(PlanState *planstate,
131 : ExecParallelEstimateContext *e);
132 : static bool ExecParallelInitializeDSM(PlanState *planstate,
133 : ExecParallelInitializeDSMContext *d);
134 : static shm_mq_handle **ExecParallelSetupTupleQueues(ParallelContext *pcxt,
135 : bool reinitialize);
136 : static bool ExecParallelReInitializeDSM(PlanState *planstate,
137 : ParallelContext *pcxt);
138 : static bool ExecParallelRetrieveInstrumentation(PlanState *planstate,
139 : SharedExecutorInstrumentation *instrumentation);
140 :
141 : /* Helper function that runs in the parallel worker. */
142 : static DestReceiver *ExecParallelGetReceiver(dsm_segment *seg, shm_toc *toc);
143 :
144 : /*
145 : * Create a serialized representation of the plan to be sent to each worker.
146 : */
147 : static char *
148 383 : ExecSerializePlan(Plan *plan, EState *estate)
149 : {
150 : PlannedStmt *pstmt;
151 : ListCell *lc;
152 :
153 : /* We can't scribble on the original plan, so make a copy. */
154 383 : plan = copyObject(plan);
155 :
156 : /*
157 : * The worker will start its own copy of the executor, and that copy will
158 : * insert a junk filter if the toplevel node has any resjunk entries. We
159 : * don't want that to happen, because while resjunk columns shouldn't be
160 : * sent back to the user, here the tuples are coming back to another
161 : * backend which may very well need them. So mutate the target list
162 : * accordingly. This is sort of a hack; there might be better ways to do
163 : * this...
164 : */
165 1046 : foreach(lc, plan->targetlist)
166 : {
167 663 : TargetEntry *tle = lfirst_node(TargetEntry, lc);
168 :
169 663 : tle->resjunk = false;
170 : }
171 :
172 : /*
173 : * Create a dummy PlannedStmt. Most of the fields don't need to be valid
174 : * for our purposes, but the worker will need at least a minimal
175 : * PlannedStmt to start the executor.
176 : */
177 383 : pstmt = makeNode(PlannedStmt);
178 383 : pstmt->commandType = CMD_SELECT;
179 383 : pstmt->queryId = pgstat_get_my_query_id();
180 383 : pstmt->planId = pgstat_get_my_plan_id();
181 383 : pstmt->hasReturning = false;
182 383 : pstmt->hasModifyingCTE = false;
183 383 : pstmt->canSetTag = true;
184 383 : pstmt->transientPlan = false;
185 383 : pstmt->dependsOnRole = false;
186 383 : pstmt->parallelModeNeeded = false;
187 383 : pstmt->planTree = plan;
188 383 : pstmt->partPruneInfos = estate->es_part_prune_infos;
189 383 : pstmt->rtable = estate->es_range_table;
190 383 : pstmt->unprunableRelids = estate->es_unpruned_relids;
191 383 : pstmt->permInfos = estate->es_rteperminfos;
192 383 : pstmt->resultRelations = NIL;
193 383 : pstmt->appendRelations = NIL;
194 383 : pstmt->planOrigin = PLAN_STMT_INTERNAL;
195 :
196 : /*
197 : * Transfer only parallel-safe subplans, leaving a NULL "hole" in the list
198 : * for unsafe ones (so that the list indexes of the safe ones are
199 : * preserved). This positively ensures that the worker won't try to run,
200 : * or even do ExecInitNode on, an unsafe subplan. That's important to
201 : * protect, eg, non-parallel-aware FDWs from getting into trouble.
202 : */
203 383 : pstmt->subplans = NIL;
204 410 : foreach(lc, estate->es_plannedstmt->subplans)
205 : {
206 27 : Plan *subplan = (Plan *) lfirst(lc);
207 :
208 27 : if (subplan && !subplan->parallel_safe)
209 6 : subplan = NULL;
210 27 : pstmt->subplans = lappend(pstmt->subplans, subplan);
211 : }
212 :
213 383 : pstmt->rewindPlanIDs = NULL;
214 383 : pstmt->rowMarks = NIL;
215 383 : pstmt->relationOids = NIL;
216 383 : pstmt->invalItems = NIL; /* workers can't replan anyway... */
217 383 : pstmt->paramExecTypes = estate->es_plannedstmt->paramExecTypes;
218 383 : pstmt->utilityStmt = NULL;
219 383 : pstmt->stmt_location = -1;
220 383 : pstmt->stmt_len = -1;
221 :
222 : /* Return serialized copy of our dummy PlannedStmt. */
223 383 : return nodeToString(pstmt);
224 : }
225 :
226 : /*
227 : * Parallel-aware plan nodes (and occasionally others) may need some state
228 : * which is shared across all parallel workers. Before we size the DSM, give
229 : * them a chance to call shm_toc_estimate_chunk or shm_toc_estimate_keys on
230 : * &pcxt->estimator.
231 : *
232 : * While we're at it, count the number of PlanState nodes in the tree, so
233 : * we know how many Instrumentation structures we need.
234 : */
235 : static bool
236 1549 : ExecParallelEstimate(PlanState *planstate, ExecParallelEstimateContext *e)
237 : {
238 1549 : if (planstate == NULL)
239 0 : return false;
240 :
241 : /* Count this node. */
242 1549 : e->nnodes++;
243 :
244 1549 : switch (nodeTag(planstate))
245 : {
246 587 : case T_SeqScanState:
247 587 : if (planstate->plan->parallel_aware)
248 468 : ExecSeqScanEstimate((SeqScanState *) planstate,
249 : e->pcxt);
250 587 : break;
251 147 : case T_IndexScanState:
252 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
253 147 : ExecIndexScanEstimate((IndexScanState *) planstate,
254 : e->pcxt);
255 147 : break;
256 29 : case T_IndexOnlyScanState:
257 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
258 29 : ExecIndexOnlyScanEstimate((IndexOnlyScanState *) planstate,
259 : e->pcxt);
260 29 : break;
261 10 : case T_BitmapIndexScanState:
262 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
263 10 : ExecBitmapIndexScanEstimate((BitmapIndexScanState *) planstate,
264 : e->pcxt);
265 10 : break;
266 0 : case T_ForeignScanState:
267 0 : if (planstate->plan->parallel_aware)
268 0 : ExecForeignScanEstimate((ForeignScanState *) planstate,
269 : e->pcxt);
270 0 : break;
271 12 : case T_TidRangeScanState:
272 12 : if (planstate->plan->parallel_aware)
273 12 : ExecTidRangeScanEstimate((TidRangeScanState *) planstate,
274 : e->pcxt);
275 12 : break;
276 93 : case T_AppendState:
277 93 : if (planstate->plan->parallel_aware)
278 69 : ExecAppendEstimate((AppendState *) planstate,
279 : e->pcxt);
280 93 : break;
281 0 : case T_CustomScanState:
282 0 : if (planstate->plan->parallel_aware)
283 0 : ExecCustomScanEstimate((CustomScanState *) planstate,
284 : e->pcxt);
285 0 : break;
286 10 : case T_BitmapHeapScanState:
287 10 : if (planstate->plan->parallel_aware)
288 9 : ExecBitmapHeapEstimate((BitmapHeapScanState *) planstate,
289 : e->pcxt);
290 10 : break;
291 99 : case T_HashJoinState:
292 99 : if (planstate->plan->parallel_aware)
293 63 : ExecHashJoinEstimate((HashJoinState *) planstate,
294 : e->pcxt);
295 99 : break;
296 99 : case T_HashState:
297 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
298 99 : ExecHashEstimate((HashState *) planstate, e->pcxt);
299 99 : break;
300 82 : case T_SortState:
301 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
302 82 : ExecSortEstimate((SortState *) planstate, e->pcxt);
303 82 : break;
304 0 : case T_IncrementalSortState:
305 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
306 0 : ExecIncrementalSortEstimate((IncrementalSortState *) planstate, e->pcxt);
307 0 : break;
308 295 : case T_AggState:
309 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
310 295 : ExecAggEstimate((AggState *) planstate, e->pcxt);
311 295 : break;
312 3 : case T_MemoizeState:
313 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
314 3 : ExecMemoizeEstimate((MemoizeState *) planstate, e->pcxt);
315 3 : break;
316 83 : default:
317 83 : break;
318 : }
319 :
320 1549 : return planstate_tree_walker(planstate, ExecParallelEstimate, e);
321 : }
322 :
323 : /*
324 : * Estimate the amount of space required to serialize the indicated parameters.
325 : */
326 : static Size
327 12 : EstimateParamExecSpace(EState *estate, Bitmapset *params)
328 : {
329 : int paramid;
330 12 : Size sz = sizeof(int);
331 :
332 12 : paramid = -1;
333 27 : while ((paramid = bms_next_member(params, paramid)) >= 0)
334 : {
335 : Oid typeOid;
336 : int16 typLen;
337 : bool typByVal;
338 : ParamExecData *prm;
339 :
340 15 : prm = &(estate->es_param_exec_vals[paramid]);
341 15 : typeOid = list_nth_oid(estate->es_plannedstmt->paramExecTypes,
342 : paramid);
343 :
344 15 : sz = add_size(sz, sizeof(int)); /* space for paramid */
345 :
346 : /* space for datum/isnull */
347 15 : if (OidIsValid(typeOid))
348 15 : get_typlenbyval(typeOid, &typLen, &typByVal);
349 : else
350 : {
351 : /* If no type OID, assume by-value, like copyParamList does. */
352 0 : typLen = sizeof(Datum);
353 0 : typByVal = true;
354 : }
355 15 : sz = add_size(sz,
356 15 : datumEstimateSpace(prm->value, prm->isnull,
357 : typByVal, typLen));
358 : }
359 12 : return sz;
360 : }
361 :
362 : /*
363 : * Serialize specified PARAM_EXEC parameters.
364 : *
365 : * We write the number of parameters first, as a 4-byte integer, and then
366 : * write details for each parameter in turn. The details for each parameter
367 : * consist of a 4-byte paramid (location of param in execution time internal
368 : * parameter array) and then the datum as serialized by datumSerialize().
369 : */
370 : static dsa_pointer
371 12 : SerializeParamExecParams(EState *estate, Bitmapset *params, dsa_area *area)
372 : {
373 : Size size;
374 : int nparams;
375 : int paramid;
376 : ParamExecData *prm;
377 : dsa_pointer handle;
378 : char *start_address;
379 :
380 : /* Allocate enough space for the current parameter values. */
381 12 : size = EstimateParamExecSpace(estate, params);
382 12 : handle = dsa_allocate(area, size);
383 12 : start_address = dsa_get_address(area, handle);
384 :
385 : /* First write the number of parameters as a 4-byte integer. */
386 12 : nparams = bms_num_members(params);
387 12 : memcpy(start_address, &nparams, sizeof(int));
388 12 : start_address += sizeof(int);
389 :
390 : /* Write details for each parameter in turn. */
391 12 : paramid = -1;
392 27 : while ((paramid = bms_next_member(params, paramid)) >= 0)
393 : {
394 : Oid typeOid;
395 : int16 typLen;
396 : bool typByVal;
397 :
398 15 : prm = &(estate->es_param_exec_vals[paramid]);
399 15 : typeOid = list_nth_oid(estate->es_plannedstmt->paramExecTypes,
400 : paramid);
401 :
402 : /* Write paramid. */
403 15 : memcpy(start_address, ¶mid, sizeof(int));
404 15 : start_address += sizeof(int);
405 :
406 : /* Write datum/isnull */
407 15 : if (OidIsValid(typeOid))
408 15 : get_typlenbyval(typeOid, &typLen, &typByVal);
409 : else
410 : {
411 : /* If no type OID, assume by-value, like copyParamList does. */
412 0 : typLen = sizeof(Datum);
413 0 : typByVal = true;
414 : }
415 15 : datumSerialize(prm->value, prm->isnull, typByVal, typLen,
416 : &start_address);
417 : }
418 :
419 12 : return handle;
420 : }
421 :
422 : /*
423 : * Restore specified PARAM_EXEC parameters.
424 : */
425 : static void
426 36 : RestoreParamExecParams(char *start_address, EState *estate)
427 : {
428 : int nparams;
429 : int i;
430 : int paramid;
431 :
432 36 : memcpy(&nparams, start_address, sizeof(int));
433 36 : start_address += sizeof(int);
434 :
435 78 : for (i = 0; i < nparams; i++)
436 : {
437 : ParamExecData *prm;
438 :
439 : /* Read paramid */
440 42 : memcpy(¶mid, start_address, sizeof(int));
441 42 : start_address += sizeof(int);
442 42 : prm = &(estate->es_param_exec_vals[paramid]);
443 :
444 : /* Read datum/isnull. */
445 42 : prm->value = datumRestore(&start_address, &prm->isnull);
446 42 : prm->execPlan = NULL;
447 : }
448 36 : }
449 :
450 : /*
451 : * Initialize the dynamic shared memory segment that will be used to control
452 : * parallel execution.
453 : */
454 : static bool
455 1549 : ExecParallelInitializeDSM(PlanState *planstate,
456 : ExecParallelInitializeDSMContext *d)
457 : {
458 1549 : if (planstate == NULL)
459 0 : return false;
460 :
461 : /* If instrumentation is enabled, initialize slot for this node. */
462 1549 : if (d->instrumentation != NULL)
463 513 : d->instrumentation->plan_node_id[d->nnodes] =
464 513 : planstate->plan->plan_node_id;
465 :
466 : /* Count this node. */
467 1549 : d->nnodes++;
468 :
469 : /*
470 : * Call initializers for DSM-using plan nodes.
471 : *
472 : * Most plan nodes won't do anything here, but plan nodes that allocated
473 : * DSM may need to initialize shared state in the DSM before parallel
474 : * workers are launched. They can allocate the space they previously
475 : * estimated using shm_toc_allocate, and add the keys they previously
476 : * estimated using shm_toc_insert, in each case targeting pcxt->toc.
477 : */
478 1549 : switch (nodeTag(planstate))
479 : {
480 587 : case T_SeqScanState:
481 587 : if (planstate->plan->parallel_aware)
482 468 : ExecSeqScanInitializeDSM((SeqScanState *) planstate,
483 : d->pcxt);
484 587 : break;
485 147 : case T_IndexScanState:
486 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
487 147 : ExecIndexScanInitializeDSM((IndexScanState *) planstate, d->pcxt);
488 147 : break;
489 29 : case T_IndexOnlyScanState:
490 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
491 29 : ExecIndexOnlyScanInitializeDSM((IndexOnlyScanState *) planstate,
492 : d->pcxt);
493 29 : break;
494 10 : case T_BitmapIndexScanState:
495 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
496 10 : ExecBitmapIndexScanInitializeDSM((BitmapIndexScanState *) planstate, d->pcxt);
497 10 : break;
498 0 : case T_ForeignScanState:
499 0 : if (planstate->plan->parallel_aware)
500 0 : ExecForeignScanInitializeDSM((ForeignScanState *) planstate,
501 : d->pcxt);
502 0 : break;
503 12 : case T_TidRangeScanState:
504 12 : if (planstate->plan->parallel_aware)
505 12 : ExecTidRangeScanInitializeDSM((TidRangeScanState *) planstate,
506 : d->pcxt);
507 12 : break;
508 93 : case T_AppendState:
509 93 : if (planstate->plan->parallel_aware)
510 69 : ExecAppendInitializeDSM((AppendState *) planstate,
511 : d->pcxt);
512 93 : break;
513 0 : case T_CustomScanState:
514 0 : if (planstate->plan->parallel_aware)
515 0 : ExecCustomScanInitializeDSM((CustomScanState *) planstate,
516 : d->pcxt);
517 0 : break;
518 10 : case T_BitmapHeapScanState:
519 10 : if (planstate->plan->parallel_aware)
520 9 : ExecBitmapHeapInitializeDSM((BitmapHeapScanState *) planstate,
521 : d->pcxt);
522 10 : break;
523 99 : case T_HashJoinState:
524 99 : if (planstate->plan->parallel_aware)
525 63 : ExecHashJoinInitializeDSM((HashJoinState *) planstate,
526 : d->pcxt);
527 99 : break;
528 99 : case T_HashState:
529 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
530 99 : ExecHashInitializeDSM((HashState *) planstate, d->pcxt);
531 99 : break;
532 82 : case T_SortState:
533 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
534 82 : ExecSortInitializeDSM((SortState *) planstate, d->pcxt);
535 82 : break;
536 0 : case T_IncrementalSortState:
537 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
538 0 : ExecIncrementalSortInitializeDSM((IncrementalSortState *) planstate, d->pcxt);
539 0 : break;
540 295 : case T_AggState:
541 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
542 295 : ExecAggInitializeDSM((AggState *) planstate, d->pcxt);
543 295 : break;
544 3 : case T_MemoizeState:
545 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
546 3 : ExecMemoizeInitializeDSM((MemoizeState *) planstate, d->pcxt);
547 3 : break;
548 83 : default:
549 83 : break;
550 : }
551 :
552 1549 : return planstate_tree_walker(planstate, ExecParallelInitializeDSM, d);
553 : }
554 :
555 : /*
556 : * It sets up the response queues for backend workers to return tuples
557 : * to the main backend and start the workers.
558 : */
559 : static shm_mq_handle **
560 512 : ExecParallelSetupTupleQueues(ParallelContext *pcxt, bool reinitialize)
561 : {
562 : shm_mq_handle **responseq;
563 : char *tqueuespace;
564 : int i;
565 :
566 : /* Skip this if no workers. */
567 512 : if (pcxt->nworkers == 0)
568 0 : return NULL;
569 :
570 : /* Allocate memory for shared memory queue handles. */
571 : responseq = (shm_mq_handle **)
572 512 : palloc(pcxt->nworkers * sizeof(shm_mq_handle *));
573 :
574 : /*
575 : * If not reinitializing, allocate space from the DSM for the queues;
576 : * otherwise, find the already allocated space.
577 : */
578 512 : if (!reinitialize)
579 : tqueuespace =
580 383 : shm_toc_allocate(pcxt->toc,
581 : mul_size(PARALLEL_TUPLE_QUEUE_SIZE,
582 383 : pcxt->nworkers));
583 : else
584 129 : tqueuespace = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_TUPLE_QUEUE, false);
585 :
586 : /* Create the queues, and become the receiver for each. */
587 1895 : for (i = 0; i < pcxt->nworkers; ++i)
588 : {
589 : shm_mq *mq;
590 :
591 1383 : mq = shm_mq_create(tqueuespace +
592 1383 : ((Size) i) * PARALLEL_TUPLE_QUEUE_SIZE,
593 : (Size) PARALLEL_TUPLE_QUEUE_SIZE);
594 :
595 1383 : shm_mq_set_receiver(mq, MyProc);
596 1383 : responseq[i] = shm_mq_attach(mq, pcxt->seg, NULL);
597 : }
598 :
599 : /* Add array of queues to shm_toc, so others can find it. */
600 512 : if (!reinitialize)
601 383 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_TUPLE_QUEUE, tqueuespace);
602 :
603 : /* Return array of handles. */
604 512 : return responseq;
605 : }
606 :
607 : /*
608 : * Sets up the required infrastructure for backend workers to perform
609 : * execution and return results to the main backend.
610 : */
611 : ParallelExecutorInfo *
612 383 : ExecInitParallelPlan(PlanState *planstate, EState *estate,
613 : Bitmapset *sendParams, int nworkers,
614 : int64 tuples_needed)
615 : {
616 : ParallelExecutorInfo *pei;
617 : ParallelContext *pcxt;
618 : ExecParallelEstimateContext e;
619 : ExecParallelInitializeDSMContext d;
620 : FixedParallelExecutorState *fpes;
621 : char *pstmt_data;
622 : char *pstmt_space;
623 : char *paramlistinfo_space;
624 : BufferUsage *bufusage_space;
625 : WalUsage *walusage_space;
626 383 : SharedExecutorInstrumentation *instrumentation = NULL;
627 383 : SharedJitInstrumentation *jit_instrumentation = NULL;
628 : int pstmt_len;
629 : int paramlistinfo_len;
630 383 : int instrumentation_len = 0;
631 383 : int jit_instrumentation_len = 0;
632 383 : int instrument_offset = 0;
633 383 : Size dsa_minsize = dsa_minimum_size();
634 : char *query_string;
635 : int query_len;
636 :
637 : /*
638 : * Force any initplan outputs that we're going to pass to workers to be
639 : * evaluated, if they weren't already.
640 : *
641 : * For simplicity, we use the EState's per-output-tuple ExprContext here.
642 : * That risks intra-query memory leakage, since we might pass through here
643 : * many times before that ExprContext gets reset; but ExecSetParamPlan
644 : * doesn't normally leak any memory in the context (see its comments), so
645 : * it doesn't seem worth complicating this function's API to pass it a
646 : * shorter-lived ExprContext. This might need to change someday.
647 : */
648 383 : ExecSetParamPlanMulti(sendParams, GetPerTupleExprContext(estate));
649 :
650 : /* Allocate object for return value. */
651 383 : pei = palloc0_object(ParallelExecutorInfo);
652 383 : pei->finished = false;
653 383 : pei->planstate = planstate;
654 :
655 : /* Fix up and serialize plan to be sent to workers. */
656 383 : pstmt_data = ExecSerializePlan(planstate->plan, estate);
657 :
658 : /* Create a parallel context. */
659 383 : pcxt = CreateParallelContext("postgres", "ParallelQueryMain", nworkers);
660 383 : pei->pcxt = pcxt;
661 :
662 : /*
663 : * Before telling the parallel context to create a dynamic shared memory
664 : * segment, we need to figure out how big it should be. Estimate space
665 : * for the various things we need to store.
666 : */
667 :
668 : /* Estimate space for fixed-size state. */
669 383 : shm_toc_estimate_chunk(&pcxt->estimator,
670 : sizeof(FixedParallelExecutorState));
671 383 : shm_toc_estimate_keys(&pcxt->estimator, 1);
672 :
673 : /* Estimate space for query text. */
674 383 : query_len = strlen(estate->es_sourceText);
675 383 : shm_toc_estimate_chunk(&pcxt->estimator, query_len + 1);
676 383 : shm_toc_estimate_keys(&pcxt->estimator, 1);
677 :
678 : /* Estimate space for serialized PlannedStmt. */
679 383 : pstmt_len = strlen(pstmt_data) + 1;
680 383 : shm_toc_estimate_chunk(&pcxt->estimator, pstmt_len);
681 383 : shm_toc_estimate_keys(&pcxt->estimator, 1);
682 :
683 : /* Estimate space for serialized ParamListInfo. */
684 383 : paramlistinfo_len = EstimateParamListSpace(estate->es_param_list_info);
685 383 : shm_toc_estimate_chunk(&pcxt->estimator, paramlistinfo_len);
686 383 : shm_toc_estimate_keys(&pcxt->estimator, 1);
687 :
688 : /*
689 : * Estimate space for BufferUsage.
690 : *
691 : * If EXPLAIN is not in use and there are no extensions loaded that care,
692 : * we could skip this. But we have no way of knowing whether anyone's
693 : * looking at pgBufferUsage, so do it unconditionally.
694 : */
695 383 : shm_toc_estimate_chunk(&pcxt->estimator,
696 : mul_size(sizeof(BufferUsage), pcxt->nworkers));
697 383 : shm_toc_estimate_keys(&pcxt->estimator, 1);
698 :
699 : /*
700 : * Same thing for WalUsage.
701 : */
702 383 : shm_toc_estimate_chunk(&pcxt->estimator,
703 : mul_size(sizeof(WalUsage), pcxt->nworkers));
704 383 : shm_toc_estimate_keys(&pcxt->estimator, 1);
705 :
706 : /* Estimate space for tuple queues. */
707 383 : shm_toc_estimate_chunk(&pcxt->estimator,
708 : mul_size(PARALLEL_TUPLE_QUEUE_SIZE, pcxt->nworkers));
709 383 : shm_toc_estimate_keys(&pcxt->estimator, 1);
710 :
711 : /*
712 : * Give parallel-aware nodes a chance to add to the estimates, and get a
713 : * count of how many PlanState nodes there are.
714 : */
715 383 : e.pcxt = pcxt;
716 383 : e.nnodes = 0;
717 383 : ExecParallelEstimate(planstate, &e);
718 :
719 : /* Estimate space for instrumentation, if required. */
720 383 : if (estate->es_instrument)
721 : {
722 90 : instrumentation_len =
723 : offsetof(SharedExecutorInstrumentation, plan_node_id) +
724 90 : sizeof(int) * e.nnodes;
725 90 : instrumentation_len = MAXALIGN(instrumentation_len);
726 90 : instrument_offset = instrumentation_len;
727 90 : instrumentation_len +=
728 90 : mul_size(sizeof(Instrumentation),
729 90 : mul_size(e.nnodes, nworkers));
730 90 : shm_toc_estimate_chunk(&pcxt->estimator, instrumentation_len);
731 90 : shm_toc_estimate_keys(&pcxt->estimator, 1);
732 :
733 : /* Estimate space for JIT instrumentation, if required. */
734 90 : if (estate->es_jit_flags != PGJIT_NONE)
735 : {
736 0 : jit_instrumentation_len =
737 0 : offsetof(SharedJitInstrumentation, jit_instr) +
738 : sizeof(JitInstrumentation) * nworkers;
739 0 : shm_toc_estimate_chunk(&pcxt->estimator, jit_instrumentation_len);
740 0 : shm_toc_estimate_keys(&pcxt->estimator, 1);
741 : }
742 : }
743 :
744 : /* Estimate space for DSA area. */
745 383 : shm_toc_estimate_chunk(&pcxt->estimator, dsa_minsize);
746 383 : shm_toc_estimate_keys(&pcxt->estimator, 1);
747 :
748 : /*
749 : * InitializeParallelDSM() passes the active snapshot to the parallel
750 : * worker, which uses it to set es_snapshot. Make sure we don't set
751 : * es_snapshot differently in the child.
752 : */
753 : Assert(GetActiveSnapshot() == estate->es_snapshot);
754 :
755 : /* Everyone's had a chance to ask for space, so now create the DSM. */
756 383 : InitializeParallelDSM(pcxt);
757 :
758 : /*
759 : * OK, now we have a dynamic shared memory segment, and it should be big
760 : * enough to store all of the data we estimated we would want to put into
761 : * it, plus whatever general stuff (not specifically executor-related) the
762 : * ParallelContext itself needs to store there. None of the space we
763 : * asked for has been allocated or initialized yet, though, so do that.
764 : */
765 :
766 : /* Store fixed-size state. */
767 383 : fpes = shm_toc_allocate(pcxt->toc, sizeof(FixedParallelExecutorState));
768 383 : fpes->tuples_needed = tuples_needed;
769 383 : fpes->param_exec = InvalidDsaPointer;
770 383 : fpes->eflags = estate->es_top_eflags;
771 383 : fpes->jit_flags = estate->es_jit_flags;
772 383 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_EXECUTOR_FIXED, fpes);
773 :
774 : /* Store query string */
775 383 : query_string = shm_toc_allocate(pcxt->toc, query_len + 1);
776 383 : memcpy(query_string, estate->es_sourceText, query_len + 1);
777 383 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_QUERY_TEXT, query_string);
778 :
779 : /* Store serialized PlannedStmt. */
780 383 : pstmt_space = shm_toc_allocate(pcxt->toc, pstmt_len);
781 383 : memcpy(pstmt_space, pstmt_data, pstmt_len);
782 383 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_PLANNEDSTMT, pstmt_space);
783 :
784 : /* Store serialized ParamListInfo. */
785 383 : paramlistinfo_space = shm_toc_allocate(pcxt->toc, paramlistinfo_len);
786 383 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_PARAMLISTINFO, paramlistinfo_space);
787 383 : SerializeParamList(estate->es_param_list_info, ¶mlistinfo_space);
788 :
789 : /* Allocate space for each worker's BufferUsage; no need to initialize. */
790 383 : bufusage_space = shm_toc_allocate(pcxt->toc,
791 383 : mul_size(sizeof(BufferUsage), pcxt->nworkers));
792 383 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_BUFFER_USAGE, bufusage_space);
793 383 : pei->buffer_usage = bufusage_space;
794 :
795 : /* Same for WalUsage. */
796 383 : walusage_space = shm_toc_allocate(pcxt->toc,
797 383 : mul_size(sizeof(WalUsage), pcxt->nworkers));
798 383 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_WAL_USAGE, walusage_space);
799 383 : pei->wal_usage = walusage_space;
800 :
801 : /* Set up the tuple queues that the workers will write into. */
802 383 : pei->tqueue = ExecParallelSetupTupleQueues(pcxt, false);
803 :
804 : /* We don't need the TupleQueueReaders yet, though. */
805 383 : pei->reader = NULL;
806 :
807 : /*
808 : * If instrumentation options were supplied, allocate space for the data.
809 : * It only gets partially initialized here; the rest happens during
810 : * ExecParallelInitializeDSM.
811 : */
812 383 : if (estate->es_instrument)
813 : {
814 : Instrumentation *instrument;
815 : int i;
816 :
817 90 : instrumentation = shm_toc_allocate(pcxt->toc, instrumentation_len);
818 90 : instrumentation->instrument_options = estate->es_instrument;
819 90 : instrumentation->instrument_offset = instrument_offset;
820 90 : instrumentation->num_workers = nworkers;
821 90 : instrumentation->num_plan_nodes = e.nnodes;
822 90 : instrument = GetInstrumentationArray(instrumentation);
823 930 : for (i = 0; i < nworkers * e.nnodes; ++i)
824 840 : InstrInit(&instrument[i], estate->es_instrument);
825 90 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_INSTRUMENTATION,
826 : instrumentation);
827 90 : pei->instrumentation = instrumentation;
828 :
829 90 : if (estate->es_jit_flags != PGJIT_NONE)
830 : {
831 0 : jit_instrumentation = shm_toc_allocate(pcxt->toc,
832 : jit_instrumentation_len);
833 0 : jit_instrumentation->num_workers = nworkers;
834 0 : memset(jit_instrumentation->jit_instr, 0,
835 : sizeof(JitInstrumentation) * nworkers);
836 0 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_JIT_INSTRUMENTATION,
837 : jit_instrumentation);
838 0 : pei->jit_instrumentation = jit_instrumentation;
839 : }
840 : }
841 :
842 : /*
843 : * Create a DSA area that can be used by the leader and all workers.
844 : * (However, if we failed to create a DSM and are using private memory
845 : * instead, then skip this.)
846 : */
847 383 : if (pcxt->seg != NULL)
848 : {
849 : char *area_space;
850 :
851 383 : area_space = shm_toc_allocate(pcxt->toc, dsa_minsize);
852 383 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_DSA, area_space);
853 383 : pei->area = dsa_create_in_place(area_space, dsa_minsize,
854 : LWTRANCHE_PARALLEL_QUERY_DSA,
855 : pcxt->seg);
856 :
857 : /*
858 : * Serialize parameters, if any, using DSA storage. We don't dare use
859 : * the main parallel query DSM for this because we might relaunch
860 : * workers after the values have changed (and thus the amount of
861 : * storage required has changed).
862 : */
863 383 : if (!bms_is_empty(sendParams))
864 : {
865 12 : pei->param_exec = SerializeParamExecParams(estate, sendParams,
866 : pei->area);
867 12 : fpes->param_exec = pei->param_exec;
868 : }
869 : }
870 :
871 : /*
872 : * Give parallel-aware nodes a chance to initialize their shared data.
873 : * This also initializes the elements of instrumentation->ps_instrument,
874 : * if it exists.
875 : */
876 383 : d.pcxt = pcxt;
877 383 : d.instrumentation = instrumentation;
878 383 : d.nnodes = 0;
879 :
880 : /* Install our DSA area while initializing the plan. */
881 383 : estate->es_query_dsa = pei->area;
882 383 : ExecParallelInitializeDSM(planstate, &d);
883 383 : estate->es_query_dsa = NULL;
884 :
885 : /*
886 : * Make sure that the world hasn't shifted under our feet. This could
887 : * probably just be an Assert(), but let's be conservative for now.
888 : */
889 383 : if (e.nnodes != d.nnodes)
890 0 : elog(ERROR, "inconsistent count of PlanState nodes");
891 :
892 : /* OK, we're ready to rock and roll. */
893 383 : return pei;
894 : }
895 :
896 : /*
897 : * Set up tuple queue readers to read the results of a parallel subplan.
898 : *
899 : * This is separate from ExecInitParallelPlan() because we can launch the
900 : * worker processes and let them start doing something before we do this.
901 : */
902 : void
903 503 : ExecParallelCreateReaders(ParallelExecutorInfo *pei)
904 : {
905 503 : int nworkers = pei->pcxt->nworkers_launched;
906 : int i;
907 :
908 : Assert(pei->reader == NULL);
909 :
910 503 : if (nworkers > 0)
911 : {
912 503 : pei->reader = (TupleQueueReader **)
913 503 : palloc(nworkers * sizeof(TupleQueueReader *));
914 :
915 1843 : for (i = 0; i < nworkers; i++)
916 : {
917 1340 : shm_mq_set_handle(pei->tqueue[i],
918 1340 : pei->pcxt->worker[i].bgwhandle);
919 1340 : pei->reader[i] = CreateTupleQueueReader(pei->tqueue[i]);
920 : }
921 : }
922 503 : }
923 :
924 : /*
925 : * Re-initialize the parallel executor shared memory state before launching
926 : * a fresh batch of workers.
927 : */
928 : void
929 129 : ExecParallelReinitialize(PlanState *planstate,
930 : ParallelExecutorInfo *pei,
931 : Bitmapset *sendParams)
932 : {
933 129 : EState *estate = planstate->state;
934 : FixedParallelExecutorState *fpes;
935 :
936 : /* Old workers must already be shut down */
937 : Assert(pei->finished);
938 :
939 : /*
940 : * Force any initplan outputs that we're going to pass to workers to be
941 : * evaluated, if they weren't already (see comments in
942 : * ExecInitParallelPlan).
943 : */
944 129 : ExecSetParamPlanMulti(sendParams, GetPerTupleExprContext(estate));
945 :
946 129 : ReinitializeParallelDSM(pei->pcxt);
947 129 : pei->tqueue = ExecParallelSetupTupleQueues(pei->pcxt, true);
948 129 : pei->reader = NULL;
949 129 : pei->finished = false;
950 :
951 129 : fpes = shm_toc_lookup(pei->pcxt->toc, PARALLEL_KEY_EXECUTOR_FIXED, false);
952 :
953 : /* Free any serialized parameters from the last round. */
954 129 : if (DsaPointerIsValid(fpes->param_exec))
955 : {
956 0 : dsa_free(pei->area, fpes->param_exec);
957 0 : fpes->param_exec = InvalidDsaPointer;
958 : }
959 :
960 : /* Serialize current parameter values if required. */
961 129 : if (!bms_is_empty(sendParams))
962 : {
963 0 : pei->param_exec = SerializeParamExecParams(estate, sendParams,
964 : pei->area);
965 0 : fpes->param_exec = pei->param_exec;
966 : }
967 :
968 : /* Traverse plan tree and let each child node reset associated state. */
969 129 : estate->es_query_dsa = pei->area;
970 129 : ExecParallelReInitializeDSM(planstate, pei->pcxt);
971 129 : estate->es_query_dsa = NULL;
972 129 : }
973 :
974 : /*
975 : * Traverse plan tree to reinitialize per-node dynamic shared memory state
976 : */
977 : static bool
978 333 : ExecParallelReInitializeDSM(PlanState *planstate,
979 : ParallelContext *pcxt)
980 : {
981 333 : if (planstate == NULL)
982 0 : return false;
983 :
984 : /*
985 : * Call reinitializers for DSM-using plan nodes.
986 : */
987 333 : switch (nodeTag(planstate))
988 : {
989 138 : case T_SeqScanState:
990 138 : if (planstate->plan->parallel_aware)
991 114 : ExecSeqScanReInitializeDSM((SeqScanState *) planstate,
992 : pcxt);
993 138 : break;
994 6 : case T_IndexScanState:
995 6 : if (planstate->plan->parallel_aware)
996 6 : ExecIndexScanReInitializeDSM((IndexScanState *) planstate,
997 : pcxt);
998 6 : break;
999 6 : case T_IndexOnlyScanState:
1000 6 : if (planstate->plan->parallel_aware)
1001 6 : ExecIndexOnlyScanReInitializeDSM((IndexOnlyScanState *) planstate,
1002 : pcxt);
1003 6 : break;
1004 0 : case T_ForeignScanState:
1005 0 : if (planstate->plan->parallel_aware)
1006 0 : ExecForeignScanReInitializeDSM((ForeignScanState *) planstate,
1007 : pcxt);
1008 0 : break;
1009 0 : case T_TidRangeScanState:
1010 0 : if (planstate->plan->parallel_aware)
1011 0 : ExecTidRangeScanReInitializeDSM((TidRangeScanState *) planstate,
1012 : pcxt);
1013 0 : break;
1014 0 : case T_AppendState:
1015 0 : if (planstate->plan->parallel_aware)
1016 0 : ExecAppendReInitializeDSM((AppendState *) planstate, pcxt);
1017 0 : break;
1018 0 : case T_CustomScanState:
1019 0 : if (planstate->plan->parallel_aware)
1020 0 : ExecCustomScanReInitializeDSM((CustomScanState *) planstate,
1021 : pcxt);
1022 0 : break;
1023 27 : case T_BitmapHeapScanState:
1024 27 : if (planstate->plan->parallel_aware)
1025 27 : ExecBitmapHeapReInitializeDSM((BitmapHeapScanState *) planstate,
1026 : pcxt);
1027 27 : break;
1028 48 : case T_HashJoinState:
1029 48 : if (planstate->plan->parallel_aware)
1030 24 : ExecHashJoinReInitializeDSM((HashJoinState *) planstate,
1031 : pcxt);
1032 48 : break;
1033 90 : case T_BitmapIndexScanState:
1034 : case T_HashState:
1035 : case T_SortState:
1036 : case T_IncrementalSortState:
1037 : case T_MemoizeState:
1038 : /* these nodes have DSM state, but no reinitialization is required */
1039 90 : break;
1040 :
1041 18 : default:
1042 18 : break;
1043 : }
1044 :
1045 333 : return planstate_tree_walker(planstate, ExecParallelReInitializeDSM, pcxt);
1046 : }
1047 :
1048 : /*
1049 : * Copy instrumentation information about this node and its descendants from
1050 : * dynamic shared memory.
1051 : */
1052 : static bool
1053 513 : ExecParallelRetrieveInstrumentation(PlanState *planstate,
1054 : SharedExecutorInstrumentation *instrumentation)
1055 : {
1056 : Instrumentation *instrument;
1057 : int i;
1058 : int n;
1059 : int ibytes;
1060 513 : int plan_node_id = planstate->plan->plan_node_id;
1061 : MemoryContext oldcontext;
1062 :
1063 : /* Find the instrumentation for this node. */
1064 2319 : for (i = 0; i < instrumentation->num_plan_nodes; ++i)
1065 2319 : if (instrumentation->plan_node_id[i] == plan_node_id)
1066 513 : break;
1067 513 : if (i >= instrumentation->num_plan_nodes)
1068 0 : elog(ERROR, "plan node %d not found", plan_node_id);
1069 :
1070 : /* Accumulate the statistics from all workers. */
1071 513 : instrument = GetInstrumentationArray(instrumentation);
1072 513 : instrument += i * instrumentation->num_workers;
1073 1353 : for (n = 0; n < instrumentation->num_workers; ++n)
1074 840 : InstrAggNode(planstate->instrument, &instrument[n]);
1075 :
1076 : /*
1077 : * Also store the per-worker detail.
1078 : *
1079 : * Worker instrumentation should be allocated in the same context as the
1080 : * regular instrumentation information, which is the per-query context.
1081 : * Switch into per-query memory context.
1082 : */
1083 513 : oldcontext = MemoryContextSwitchTo(planstate->state->es_query_cxt);
1084 513 : ibytes = mul_size(instrumentation->num_workers, sizeof(Instrumentation));
1085 513 : planstate->worker_instrument =
1086 513 : palloc(ibytes + offsetof(WorkerInstrumentation, instrument));
1087 513 : MemoryContextSwitchTo(oldcontext);
1088 :
1089 513 : planstate->worker_instrument->num_workers = instrumentation->num_workers;
1090 513 : memcpy(&planstate->worker_instrument->instrument, instrument, ibytes);
1091 :
1092 : /* Perform any node-type-specific work that needs to be done. */
1093 513 : switch (nodeTag(planstate))
1094 : {
1095 135 : case T_IndexScanState:
1096 135 : ExecIndexScanRetrieveInstrumentation((IndexScanState *) planstate);
1097 135 : break;
1098 0 : case T_IndexOnlyScanState:
1099 0 : ExecIndexOnlyScanRetrieveInstrumentation((IndexOnlyScanState *) planstate);
1100 0 : break;
1101 0 : case T_BitmapIndexScanState:
1102 0 : ExecBitmapIndexScanRetrieveInstrumentation((BitmapIndexScanState *) planstate);
1103 0 : break;
1104 6 : case T_SortState:
1105 6 : ExecSortRetrieveInstrumentation((SortState *) planstate);
1106 6 : break;
1107 0 : case T_IncrementalSortState:
1108 0 : ExecIncrementalSortRetrieveInstrumentation((IncrementalSortState *) planstate);
1109 0 : break;
1110 42 : case T_HashState:
1111 42 : ExecHashRetrieveInstrumentation((HashState *) planstate);
1112 42 : break;
1113 51 : case T_AggState:
1114 51 : ExecAggRetrieveInstrumentation((AggState *) planstate);
1115 51 : break;
1116 0 : case T_MemoizeState:
1117 0 : ExecMemoizeRetrieveInstrumentation((MemoizeState *) planstate);
1118 0 : break;
1119 0 : case T_BitmapHeapScanState:
1120 0 : ExecBitmapHeapRetrieveInstrumentation((BitmapHeapScanState *) planstate);
1121 0 : break;
1122 279 : default:
1123 279 : break;
1124 : }
1125 :
1126 513 : return planstate_tree_walker(planstate, ExecParallelRetrieveInstrumentation,
1127 : instrumentation);
1128 : }
1129 :
1130 : /*
1131 : * Add up the workers' JIT instrumentation from dynamic shared memory.
1132 : */
1133 : static void
1134 0 : ExecParallelRetrieveJitInstrumentation(PlanState *planstate,
1135 : SharedJitInstrumentation *shared_jit)
1136 : {
1137 : JitInstrumentation *combined;
1138 : int ibytes;
1139 :
1140 : int n;
1141 :
1142 : /*
1143 : * Accumulate worker JIT instrumentation into the combined JIT
1144 : * instrumentation, allocating it if required.
1145 : */
1146 0 : if (!planstate->state->es_jit_worker_instr)
1147 0 : planstate->state->es_jit_worker_instr =
1148 0 : MemoryContextAllocZero(planstate->state->es_query_cxt, sizeof(JitInstrumentation));
1149 0 : combined = planstate->state->es_jit_worker_instr;
1150 :
1151 : /* Accumulate all the workers' instrumentations. */
1152 0 : for (n = 0; n < shared_jit->num_workers; ++n)
1153 0 : InstrJitAgg(combined, &shared_jit->jit_instr[n]);
1154 :
1155 : /*
1156 : * Store the per-worker detail.
1157 : *
1158 : * Similar to ExecParallelRetrieveInstrumentation(), allocate the
1159 : * instrumentation in per-query context.
1160 : */
1161 0 : ibytes = offsetof(SharedJitInstrumentation, jit_instr)
1162 0 : + mul_size(shared_jit->num_workers, sizeof(JitInstrumentation));
1163 0 : planstate->worker_jit_instrument =
1164 0 : MemoryContextAlloc(planstate->state->es_query_cxt, ibytes);
1165 :
1166 0 : memcpy(planstate->worker_jit_instrument, shared_jit, ibytes);
1167 0 : }
1168 :
1169 : /*
1170 : * Finish parallel execution. We wait for parallel workers to finish, and
1171 : * accumulate their buffer/WAL usage.
1172 : */
1173 : void
1174 922 : ExecParallelFinish(ParallelExecutorInfo *pei)
1175 : {
1176 922 : int nworkers = pei->pcxt->nworkers_launched;
1177 : int i;
1178 :
1179 : /* Make this be a no-op if called twice in a row. */
1180 922 : if (pei->finished)
1181 416 : return;
1182 :
1183 : /*
1184 : * Detach from tuple queues ASAP, so that any still-active workers will
1185 : * notice that no further results are wanted.
1186 : */
1187 506 : if (pei->tqueue != NULL)
1188 : {
1189 1840 : for (i = 0; i < nworkers; i++)
1190 1334 : shm_mq_detach(pei->tqueue[i]);
1191 506 : pfree(pei->tqueue);
1192 506 : pei->tqueue = NULL;
1193 : }
1194 :
1195 : /*
1196 : * While we're waiting for the workers to finish, let's get rid of the
1197 : * tuple queue readers. (Any other local cleanup could be done here too.)
1198 : */
1199 506 : if (pei->reader != NULL)
1200 : {
1201 1831 : for (i = 0; i < nworkers; i++)
1202 1334 : DestroyTupleQueueReader(pei->reader[i]);
1203 497 : pfree(pei->reader);
1204 497 : pei->reader = NULL;
1205 : }
1206 :
1207 : /* Now wait for the workers to finish. */
1208 506 : WaitForParallelWorkersToFinish(pei->pcxt);
1209 :
1210 : /*
1211 : * Next, accumulate buffer/WAL usage. (This must wait for the workers to
1212 : * finish, or we might get incomplete data.)
1213 : */
1214 1840 : for (i = 0; i < nworkers; i++)
1215 1334 : InstrAccumParallelQuery(&pei->buffer_usage[i], &pei->wal_usage[i]);
1216 :
1217 506 : pei->finished = true;
1218 : }
1219 :
1220 : /*
1221 : * Accumulate instrumentation, and then clean up whatever ParallelExecutorInfo
1222 : * resources still exist after ExecParallelFinish. We separate these
1223 : * routines because someone might want to examine the contents of the DSM
1224 : * after ExecParallelFinish and before calling this routine.
1225 : */
1226 : void
1227 377 : ExecParallelCleanup(ParallelExecutorInfo *pei)
1228 : {
1229 : /* Accumulate instrumentation, if any. */
1230 377 : if (pei->instrumentation)
1231 90 : ExecParallelRetrieveInstrumentation(pei->planstate,
1232 : pei->instrumentation);
1233 :
1234 : /* Accumulate JIT instrumentation, if any. */
1235 377 : if (pei->jit_instrumentation)
1236 0 : ExecParallelRetrieveJitInstrumentation(pei->planstate,
1237 0 : pei->jit_instrumentation);
1238 :
1239 : /* Free any serialized parameters. */
1240 377 : if (DsaPointerIsValid(pei->param_exec))
1241 : {
1242 12 : dsa_free(pei->area, pei->param_exec);
1243 12 : pei->param_exec = InvalidDsaPointer;
1244 : }
1245 377 : if (pei->area != NULL)
1246 : {
1247 377 : dsa_detach(pei->area);
1248 377 : pei->area = NULL;
1249 : }
1250 377 : if (pei->pcxt != NULL)
1251 : {
1252 377 : DestroyParallelContext(pei->pcxt);
1253 377 : pei->pcxt = NULL;
1254 : }
1255 377 : pfree(pei);
1256 377 : }
1257 :
1258 : /*
1259 : * Create a DestReceiver to write tuples we produce to the shm_mq designated
1260 : * for that purpose.
1261 : */
1262 : static DestReceiver *
1263 1340 : ExecParallelGetReceiver(dsm_segment *seg, shm_toc *toc)
1264 : {
1265 : char *mqspace;
1266 : shm_mq *mq;
1267 :
1268 1340 : mqspace = shm_toc_lookup(toc, PARALLEL_KEY_TUPLE_QUEUE, false);
1269 1340 : mqspace += ParallelWorkerNumber * PARALLEL_TUPLE_QUEUE_SIZE;
1270 1340 : mq = (shm_mq *) mqspace;
1271 1340 : shm_mq_set_sender(mq, MyProc);
1272 1340 : return CreateTupleQueueDestReceiver(shm_mq_attach(mq, seg, NULL));
1273 : }
1274 :
1275 : /*
1276 : * Create a QueryDesc for the PlannedStmt we are to execute, and return it.
1277 : */
1278 : static QueryDesc *
1279 1340 : ExecParallelGetQueryDesc(shm_toc *toc, DestReceiver *receiver,
1280 : int instrument_options)
1281 : {
1282 : char *pstmtspace;
1283 : char *paramspace;
1284 : PlannedStmt *pstmt;
1285 : ParamListInfo paramLI;
1286 : char *queryString;
1287 :
1288 : /* Get the query string from shared memory */
1289 1340 : queryString = shm_toc_lookup(toc, PARALLEL_KEY_QUERY_TEXT, false);
1290 :
1291 : /* Reconstruct leader-supplied PlannedStmt. */
1292 1340 : pstmtspace = shm_toc_lookup(toc, PARALLEL_KEY_PLANNEDSTMT, false);
1293 1340 : pstmt = (PlannedStmt *) stringToNode(pstmtspace);
1294 :
1295 : /* Reconstruct ParamListInfo. */
1296 1340 : paramspace = shm_toc_lookup(toc, PARALLEL_KEY_PARAMLISTINFO, false);
1297 1340 : paramLI = RestoreParamList(¶mspace);
1298 :
1299 : /* Create a QueryDesc for the query. */
1300 1340 : return CreateQueryDesc(pstmt,
1301 : queryString,
1302 : GetActiveSnapshot(), InvalidSnapshot,
1303 : receiver, paramLI, NULL, instrument_options);
1304 : }
1305 :
1306 : /*
1307 : * Copy instrumentation information from this node and its descendants into
1308 : * dynamic shared memory, so that the parallel leader can retrieve it.
1309 : */
1310 : static bool
1311 1188 : ExecParallelReportInstrumentation(PlanState *planstate,
1312 : SharedExecutorInstrumentation *instrumentation)
1313 : {
1314 : int i;
1315 1188 : int plan_node_id = planstate->plan->plan_node_id;
1316 : Instrumentation *instrument;
1317 :
1318 1188 : InstrEndLoop(planstate->instrument);
1319 :
1320 : /*
1321 : * If we shuffled the plan_node_id values in ps_instrument into sorted
1322 : * order, we could use binary search here. This might matter someday if
1323 : * we're pushing down sufficiently large plan trees. For now, do it the
1324 : * slow, dumb way.
1325 : */
1326 3906 : for (i = 0; i < instrumentation->num_plan_nodes; ++i)
1327 3906 : if (instrumentation->plan_node_id[i] == plan_node_id)
1328 1188 : break;
1329 1188 : if (i >= instrumentation->num_plan_nodes)
1330 0 : elog(ERROR, "plan node %d not found", plan_node_id);
1331 :
1332 : /*
1333 : * Add our statistics to the per-node, per-worker totals. It's possible
1334 : * that this could happen more than once if we relaunched workers.
1335 : */
1336 1188 : instrument = GetInstrumentationArray(instrumentation);
1337 1188 : instrument += i * instrumentation->num_workers;
1338 : Assert(IsParallelWorker());
1339 : Assert(ParallelWorkerNumber < instrumentation->num_workers);
1340 1188 : InstrAggNode(&instrument[ParallelWorkerNumber], planstate->instrument);
1341 :
1342 1188 : return planstate_tree_walker(planstate, ExecParallelReportInstrumentation,
1343 : instrumentation);
1344 : }
1345 :
1346 : /*
1347 : * Initialize the PlanState and its descendants with the information
1348 : * retrieved from shared memory. This has to be done once the PlanState
1349 : * is allocated and initialized by executor; that is, after ExecutorStart().
1350 : */
1351 : static bool
1352 4258 : ExecParallelInitializeWorker(PlanState *planstate, ParallelWorkerContext *pwcxt)
1353 : {
1354 4258 : if (planstate == NULL)
1355 0 : return false;
1356 :
1357 4258 : switch (nodeTag(planstate))
1358 : {
1359 1691 : case T_SeqScanState:
1360 1691 : if (planstate->plan->parallel_aware)
1361 1377 : ExecSeqScanInitializeWorker((SeqScanState *) planstate, pwcxt);
1362 1691 : break;
1363 198 : case T_IndexScanState:
1364 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
1365 198 : ExecIndexScanInitializeWorker((IndexScanState *) planstate, pwcxt);
1366 198 : break;
1367 121 : case T_IndexOnlyScanState:
1368 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
1369 121 : ExecIndexOnlyScanInitializeWorker((IndexOnlyScanState *) planstate,
1370 : pwcxt);
1371 121 : break;
1372 135 : case T_BitmapIndexScanState:
1373 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
1374 135 : ExecBitmapIndexScanInitializeWorker((BitmapIndexScanState *) planstate,
1375 : pwcxt);
1376 135 : break;
1377 0 : case T_ForeignScanState:
1378 0 : if (planstate->plan->parallel_aware)
1379 0 : ExecForeignScanInitializeWorker((ForeignScanState *) planstate,
1380 : pwcxt);
1381 0 : break;
1382 48 : case T_TidRangeScanState:
1383 48 : if (planstate->plan->parallel_aware)
1384 48 : ExecTidRangeScanInitializeWorker((TidRangeScanState *) planstate,
1385 : pwcxt);
1386 48 : break;
1387 187 : case T_AppendState:
1388 187 : if (planstate->plan->parallel_aware)
1389 157 : ExecAppendInitializeWorker((AppendState *) planstate, pwcxt);
1390 187 : break;
1391 0 : case T_CustomScanState:
1392 0 : if (planstate->plan->parallel_aware)
1393 0 : ExecCustomScanInitializeWorker((CustomScanState *) planstate,
1394 : pwcxt);
1395 0 : break;
1396 135 : case T_BitmapHeapScanState:
1397 135 : if (planstate->plan->parallel_aware)
1398 134 : ExecBitmapHeapInitializeWorker((BitmapHeapScanState *) planstate,
1399 : pwcxt);
1400 135 : break;
1401 279 : case T_HashJoinState:
1402 279 : if (planstate->plan->parallel_aware)
1403 159 : ExecHashJoinInitializeWorker((HashJoinState *) planstate,
1404 : pwcxt);
1405 279 : break;
1406 279 : case T_HashState:
1407 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
1408 279 : ExecHashInitializeWorker((HashState *) planstate, pwcxt);
1409 279 : break;
1410 236 : case T_SortState:
1411 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
1412 236 : ExecSortInitializeWorker((SortState *) planstate, pwcxt);
1413 236 : break;
1414 0 : case T_IncrementalSortState:
1415 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
1416 0 : ExecIncrementalSortInitializeWorker((IncrementalSortState *) planstate,
1417 : pwcxt);
1418 0 : break;
1419 830 : case T_AggState:
1420 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
1421 830 : ExecAggInitializeWorker((AggState *) planstate, pwcxt);
1422 830 : break;
1423 6 : case T_MemoizeState:
1424 : /* even when not parallel-aware, for EXPLAIN ANALYZE */
1425 6 : ExecMemoizeInitializeWorker((MemoizeState *) planstate, pwcxt);
1426 6 : break;
1427 113 : default:
1428 113 : break;
1429 : }
1430 :
1431 4258 : return planstate_tree_walker(planstate, ExecParallelInitializeWorker,
1432 : pwcxt);
1433 : }
1434 :
1435 : /*
1436 : * Main entrypoint for parallel query worker processes.
1437 : *
1438 : * We reach this function from ParallelWorkerMain, so the setup necessary to
1439 : * create a sensible parallel environment has already been done;
1440 : * ParallelWorkerMain worries about stuff like the transaction state, combo
1441 : * CID mappings, and GUC values, so we don't need to deal with any of that
1442 : * here.
1443 : *
1444 : * Our job is to deal with concerns specific to the executor. The parallel
1445 : * group leader will have stored a serialized PlannedStmt, and it's our job
1446 : * to execute that plan and write the resulting tuples to the appropriate
1447 : * tuple queue. Various bits of supporting information that we need in order
1448 : * to do this are also stored in the dsm_segment and can be accessed through
1449 : * the shm_toc.
1450 : */
1451 : void
1452 1340 : ParallelQueryMain(dsm_segment *seg, shm_toc *toc)
1453 : {
1454 : FixedParallelExecutorState *fpes;
1455 : BufferUsage *buffer_usage;
1456 : WalUsage *wal_usage;
1457 : DestReceiver *receiver;
1458 : QueryDesc *queryDesc;
1459 : SharedExecutorInstrumentation *instrumentation;
1460 : SharedJitInstrumentation *jit_instrumentation;
1461 1340 : int instrument_options = 0;
1462 : void *area_space;
1463 : dsa_area *area;
1464 : ParallelWorkerContext pwcxt;
1465 :
1466 : /* Get fixed-size state. */
1467 1340 : fpes = shm_toc_lookup(toc, PARALLEL_KEY_EXECUTOR_FIXED, false);
1468 :
1469 : /* Set up DestReceiver, SharedExecutorInstrumentation, and QueryDesc. */
1470 1340 : receiver = ExecParallelGetReceiver(seg, toc);
1471 1340 : instrumentation = shm_toc_lookup(toc, PARALLEL_KEY_INSTRUMENTATION, true);
1472 1340 : if (instrumentation != NULL)
1473 363 : instrument_options = instrumentation->instrument_options;
1474 1340 : jit_instrumentation = shm_toc_lookup(toc, PARALLEL_KEY_JIT_INSTRUMENTATION,
1475 : true);
1476 1340 : queryDesc = ExecParallelGetQueryDesc(toc, receiver, instrument_options);
1477 :
1478 : /* Setting debug_query_string for individual workers */
1479 1340 : debug_query_string = queryDesc->sourceText;
1480 :
1481 : /* Report workers' query for monitoring purposes */
1482 1340 : pgstat_report_activity(STATE_RUNNING, debug_query_string);
1483 :
1484 : /* Attach to the dynamic shared memory area. */
1485 1340 : area_space = shm_toc_lookup(toc, PARALLEL_KEY_DSA, false);
1486 1340 : area = dsa_attach_in_place(area_space, seg);
1487 :
1488 : /* Start up the executor */
1489 1340 : queryDesc->plannedstmt->jitFlags = fpes->jit_flags;
1490 1340 : ExecutorStart(queryDesc, fpes->eflags);
1491 :
1492 : /* Special executor initialization steps for parallel workers */
1493 1340 : queryDesc->planstate->state->es_query_dsa = area;
1494 1340 : if (DsaPointerIsValid(fpes->param_exec))
1495 : {
1496 : char *paramexec_space;
1497 :
1498 36 : paramexec_space = dsa_get_address(area, fpes->param_exec);
1499 36 : RestoreParamExecParams(paramexec_space, queryDesc->estate);
1500 : }
1501 1340 : pwcxt.toc = toc;
1502 1340 : pwcxt.seg = seg;
1503 1340 : ExecParallelInitializeWorker(queryDesc->planstate, &pwcxt);
1504 :
1505 : /* Pass down any tuple bound */
1506 1340 : ExecSetTupleBound(fpes->tuples_needed, queryDesc->planstate);
1507 :
1508 : /*
1509 : * Prepare to track buffer/WAL usage during query execution.
1510 : *
1511 : * We do this after starting up the executor to match what happens in the
1512 : * leader, which also doesn't count buffer accesses and WAL activity that
1513 : * occur during executor startup.
1514 : */
1515 1340 : InstrStartParallelQuery();
1516 :
1517 : /*
1518 : * Run the plan. If we specified a tuple bound, be careful not to demand
1519 : * more tuples than that.
1520 : */
1521 1340 : ExecutorRun(queryDesc,
1522 : ForwardScanDirection,
1523 1340 : fpes->tuples_needed < 0 ? (int64) 0 : fpes->tuples_needed);
1524 :
1525 : /* Shut down the executor */
1526 1334 : ExecutorFinish(queryDesc);
1527 :
1528 : /* Report buffer/WAL usage during parallel execution. */
1529 1334 : buffer_usage = shm_toc_lookup(toc, PARALLEL_KEY_BUFFER_USAGE, false);
1530 1334 : wal_usage = shm_toc_lookup(toc, PARALLEL_KEY_WAL_USAGE, false);
1531 1334 : InstrEndParallelQuery(&buffer_usage[ParallelWorkerNumber],
1532 1334 : &wal_usage[ParallelWorkerNumber]);
1533 :
1534 : /* Report instrumentation data if any instrumentation options are set. */
1535 1334 : if (instrumentation != NULL)
1536 363 : ExecParallelReportInstrumentation(queryDesc->planstate,
1537 : instrumentation);
1538 :
1539 : /* Report JIT instrumentation data if any */
1540 1334 : if (queryDesc->estate->es_jit && jit_instrumentation != NULL)
1541 : {
1542 : Assert(ParallelWorkerNumber < jit_instrumentation->num_workers);
1543 0 : jit_instrumentation->jit_instr[ParallelWorkerNumber] =
1544 0 : queryDesc->estate->es_jit->instr;
1545 : }
1546 :
1547 : /* Must do this after capturing instrumentation. */
1548 1334 : ExecutorEnd(queryDesc);
1549 :
1550 : /* Cleanup. */
1551 1334 : dsa_detach(area);
1552 1334 : FreeQueryDesc(queryDesc);
1553 1334 : receiver->rDestroy(receiver);
1554 1334 : }
|
