1 /*-------------------------------------------------------------------------
2 *
3 * execParallel.c
4 * Support routines for parallel execution.
5 *
6 * Portions Copyright (c) 1996-2019, 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 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/nodeAppend.h"
29 #include "executor/nodeBitmapHeapscan.h"
30 #include "executor/nodeCustom.h"
31 #include "executor/nodeForeignscan.h"
32 #include "executor/nodeHash.h"
33 #include "executor/nodeHashjoin.h"
34 #include "executor/nodeIndexscan.h"
35 #include "executor/nodeIndexonlyscan.h"
36 #include "executor/nodeSeqscan.h"
37 #include "executor/nodeSort.h"
38 #include "executor/nodeSubplan.h"
39 #include "executor/tqueue.h"
40 #include "jit/jit.h"
41 #include "nodes/nodeFuncs.h"
42 #include "storage/spin.h"
43 #include "tcop/tcopprot.h"
44 #include "utils/datum.h"
45 #include "utils/dsa.h"
46 #include "utils/lsyscache.h"
47 #include "utils/memutils.h"
48 #include "utils/snapmgr.h"
49 #include "pgstat.h"
50
51 /*
52 * Magic numbers for parallel executor communication. We use constants
53 * greater than any 32-bit integer here so that values < 2^32 can be used
54 * by individual parallel nodes to store their own state.
55 */
56 #define PARALLEL_KEY_EXECUTOR_FIXED UINT64CONST(0xE000000000000001)
57 #define PARALLEL_KEY_PLANNEDSTMT UINT64CONST(0xE000000000000002)
58 #define PARALLEL_KEY_PARAMLISTINFO UINT64CONST(0xE000000000000003)
59 #define PARALLEL_KEY_BUFFER_USAGE UINT64CONST(0xE000000000000004)
60 #define PARALLEL_KEY_TUPLE_QUEUE UINT64CONST(0xE000000000000005)
61 #define PARALLEL_KEY_INSTRUMENTATION UINT64CONST(0xE000000000000006)
62 #define PARALLEL_KEY_DSA UINT64CONST(0xE000000000000007)
63 #define PARALLEL_KEY_QUERY_TEXT UINT64CONST(0xE000000000000008)
64 #define PARALLEL_KEY_JIT_INSTRUMENTATION UINT64CONST(0xE000000000000009)
65
66 #define PARALLEL_TUPLE_QUEUE_SIZE 65536
67
68 /*
69 * Fixed-size random stuff that we need to pass to parallel workers.
70 */
71 typedef struct FixedParallelExecutorState
72 {
73 int64 tuples_needed; /* tuple bound, see ExecSetTupleBound */
74 dsa_pointer param_exec;
75 int eflags;
76 int jit_flags;
77 } FixedParallelExecutorState;
78
79 /*
80 * DSM structure for accumulating per-PlanState instrumentation.
81 *
82 * instrument_options: Same meaning here as in instrument.c.
83 *
84 * instrument_offset: Offset, relative to the start of this structure,
85 * of the first Instrumentation object. This will depend on the length of
86 * the plan_node_id array.
87 *
88 * num_workers: Number of workers.
89 *
90 * num_plan_nodes: Number of plan nodes.
91 *
92 * plan_node_id: Array of plan nodes for which we are gathering instrumentation
93 * from parallel workers. The length of this array is given by num_plan_nodes.
94 */
95 struct SharedExecutorInstrumentation
96 {
97 int instrument_options;
98 int instrument_offset;
99 int num_workers;
100 int num_plan_nodes;
101 int plan_node_id[FLEXIBLE_ARRAY_MEMBER];
102 /* array of num_plan_nodes * num_workers Instrumentation objects follows */
103 };
104 #define GetInstrumentationArray(sei) \
105 (AssertVariableIsOfTypeMacro(sei, SharedExecutorInstrumentation *), \
106 (Instrumentation *) (((char *) sei) + sei->instrument_offset))
107
108 /* Context object for ExecParallelEstimate. */
109 typedef struct ExecParallelEstimateContext
110 {
111 ParallelContext *pcxt;
112 int nnodes;
113 } ExecParallelEstimateContext;
114
115 /* Context object for ExecParallelInitializeDSM. */
116 typedef struct ExecParallelInitializeDSMContext
117 {
118 ParallelContext *pcxt;
119 SharedExecutorInstrumentation *instrumentation;
120 int nnodes;
121 } ExecParallelInitializeDSMContext;
122
123 /* Helper functions that run in the parallel leader. */
124 static char *ExecSerializePlan(Plan *plan, EState *estate);
125 static bool ExecParallelEstimate(PlanState *node,
126 ExecParallelEstimateContext *e);
127 static bool ExecParallelInitializeDSM(PlanState *node,
128 ExecParallelInitializeDSMContext *d);
129 static shm_mq_handle **ExecParallelSetupTupleQueues(ParallelContext *pcxt,
130 bool reinitialize);
131 static bool ExecParallelReInitializeDSM(PlanState *planstate,
132 ParallelContext *pcxt);
133 static bool ExecParallelRetrieveInstrumentation(PlanState *planstate,
134 SharedExecutorInstrumentation *instrumentation);
135
136 /* Helper function that runs in the parallel worker. */
137 static DestReceiver *ExecParallelGetReceiver(dsm_segment *seg, shm_toc *toc);
138
139 /*
140 * Create a serialized representation of the plan to be sent to each worker.
141 */
142 static char *
ExecSerializePlan(Plan * plan,EState * estate)143 ExecSerializePlan(Plan *plan, EState *estate)
144 {
145 PlannedStmt *pstmt;
146 ListCell *lc;
147
148 /* We can't scribble on the original plan, so make a copy. */
149 plan = copyObject(plan);
150
151 /*
152 * The worker will start its own copy of the executor, and that copy will
153 * insert a junk filter if the toplevel node has any resjunk entries. We
154 * don't want that to happen, because while resjunk columns shouldn't be
155 * sent back to the user, here the tuples are coming back to another
156 * backend which may very well need them. So mutate the target list
157 * accordingly. This is sort of a hack; there might be better ways to do
158 * this...
159 */
160 foreach(lc, plan->targetlist)
161 {
162 TargetEntry *tle = lfirst_node(TargetEntry, lc);
163
164 tle->resjunk = false;
165 }
166
167 /*
168 * Create a dummy PlannedStmt. Most of the fields don't need to be valid
169 * for our purposes, but the worker will need at least a minimal
170 * PlannedStmt to start the executor.
171 */
172 pstmt = makeNode(PlannedStmt);
173 pstmt->commandType = CMD_SELECT;
174 pstmt->queryId = UINT64CONST(0);
175 pstmt->hasReturning = false;
176 pstmt->hasModifyingCTE = false;
177 pstmt->canSetTag = true;
178 pstmt->transientPlan = false;
179 pstmt->dependsOnRole = false;
180 pstmt->parallelModeNeeded = false;
181 pstmt->planTree = plan;
182 pstmt->rtable = estate->es_range_table;
183 pstmt->resultRelations = NIL;
184
185 /*
186 * Transfer only parallel-safe subplans, leaving a NULL "hole" in the list
187 * for unsafe ones (so that the list indexes of the safe ones are
188 * preserved). This positively ensures that the worker won't try to run,
189 * or even do ExecInitNode on, an unsafe subplan. That's important to
190 * protect, eg, non-parallel-aware FDWs from getting into trouble.
191 */
192 pstmt->subplans = NIL;
193 foreach(lc, estate->es_plannedstmt->subplans)
194 {
195 Plan *subplan = (Plan *) lfirst(lc);
196
197 if (subplan && !subplan->parallel_safe)
198 subplan = NULL;
199 pstmt->subplans = lappend(pstmt->subplans, subplan);
200 }
201
202 pstmt->rewindPlanIDs = NULL;
203 pstmt->rowMarks = NIL;
204 pstmt->relationOids = NIL;
205 pstmt->invalItems = NIL; /* workers can't replan anyway... */
206 pstmt->paramExecTypes = estate->es_plannedstmt->paramExecTypes;
207 pstmt->utilityStmt = NULL;
208 pstmt->stmt_location = -1;
209 pstmt->stmt_len = -1;
210
211 /* Return serialized copy of our dummy PlannedStmt. */
212 return nodeToString(pstmt);
213 }
214
215 /*
216 * Parallel-aware plan nodes (and occasionally others) may need some state
217 * which is shared across all parallel workers. Before we size the DSM, give
218 * them a chance to call shm_toc_estimate_chunk or shm_toc_estimate_keys on
219 * &pcxt->estimator.
220 *
221 * While we're at it, count the number of PlanState nodes in the tree, so
222 * we know how many Instrumentation structures we need.
223 */
224 static bool
ExecParallelEstimate(PlanState * planstate,ExecParallelEstimateContext * e)225 ExecParallelEstimate(PlanState *planstate, ExecParallelEstimateContext *e)
226 {
227 if (planstate == NULL)
228 return false;
229
230 /* Count this node. */
231 e->nnodes++;
232
233 switch (nodeTag(planstate))
234 {
235 case T_SeqScanState:
236 if (planstate->plan->parallel_aware)
237 ExecSeqScanEstimate((SeqScanState *) planstate,
238 e->pcxt);
239 break;
240 case T_IndexScanState:
241 if (planstate->plan->parallel_aware)
242 ExecIndexScanEstimate((IndexScanState *) planstate,
243 e->pcxt);
244 break;
245 case T_IndexOnlyScanState:
246 if (planstate->plan->parallel_aware)
247 ExecIndexOnlyScanEstimate((IndexOnlyScanState *) planstate,
248 e->pcxt);
249 break;
250 case T_ForeignScanState:
251 if (planstate->plan->parallel_aware)
252 ExecForeignScanEstimate((ForeignScanState *) planstate,
253 e->pcxt);
254 break;
255 case T_AppendState:
256 if (planstate->plan->parallel_aware)
257 ExecAppendEstimate((AppendState *) planstate,
258 e->pcxt);
259 break;
260 case T_CustomScanState:
261 if (planstate->plan->parallel_aware)
262 ExecCustomScanEstimate((CustomScanState *) planstate,
263 e->pcxt);
264 break;
265 case T_BitmapHeapScanState:
266 if (planstate->plan->parallel_aware)
267 ExecBitmapHeapEstimate((BitmapHeapScanState *) planstate,
268 e->pcxt);
269 break;
270 case T_HashJoinState:
271 if (planstate->plan->parallel_aware)
272 ExecHashJoinEstimate((HashJoinState *) planstate,
273 e->pcxt);
274 break;
275 case T_HashState:
276 /* even when not parallel-aware, for EXPLAIN ANALYZE */
277 ExecHashEstimate((HashState *) planstate, e->pcxt);
278 break;
279 case T_SortState:
280 /* even when not parallel-aware, for EXPLAIN ANALYZE */
281 ExecSortEstimate((SortState *) planstate, e->pcxt);
282 break;
283
284 default:
285 break;
286 }
287
288 return planstate_tree_walker(planstate, ExecParallelEstimate, e);
289 }
290
291 /*
292 * Estimate the amount of space required to serialize the indicated parameters.
293 */
294 static Size
EstimateParamExecSpace(EState * estate,Bitmapset * params)295 EstimateParamExecSpace(EState *estate, Bitmapset *params)
296 {
297 int paramid;
298 Size sz = sizeof(int);
299
300 paramid = -1;
301 while ((paramid = bms_next_member(params, paramid)) >= 0)
302 {
303 Oid typeOid;
304 int16 typLen;
305 bool typByVal;
306 ParamExecData *prm;
307
308 prm = &(estate->es_param_exec_vals[paramid]);
309 typeOid = list_nth_oid(estate->es_plannedstmt->paramExecTypes,
310 paramid);
311
312 sz = add_size(sz, sizeof(int)); /* space for paramid */
313
314 /* space for datum/isnull */
315 if (OidIsValid(typeOid))
316 get_typlenbyval(typeOid, &typLen, &typByVal);
317 else
318 {
319 /* If no type OID, assume by-value, like copyParamList does. */
320 typLen = sizeof(Datum);
321 typByVal = true;
322 }
323 sz = add_size(sz,
324 datumEstimateSpace(prm->value, prm->isnull,
325 typByVal, typLen));
326 }
327 return sz;
328 }
329
330 /*
331 * Serialize specified PARAM_EXEC parameters.
332 *
333 * We write the number of parameters first, as a 4-byte integer, and then
334 * write details for each parameter in turn. The details for each parameter
335 * consist of a 4-byte paramid (location of param in execution time internal
336 * parameter array) and then the datum as serialized by datumSerialize().
337 */
338 static dsa_pointer
SerializeParamExecParams(EState * estate,Bitmapset * params,dsa_area * area)339 SerializeParamExecParams(EState *estate, Bitmapset *params, dsa_area *area)
340 {
341 Size size;
342 int nparams;
343 int paramid;
344 ParamExecData *prm;
345 dsa_pointer handle;
346 char *start_address;
347
348 /* Allocate enough space for the current parameter values. */
349 size = EstimateParamExecSpace(estate, params);
350 handle = dsa_allocate(area, size);
351 start_address = dsa_get_address(area, handle);
352
353 /* First write the number of parameters as a 4-byte integer. */
354 nparams = bms_num_members(params);
355 memcpy(start_address, &nparams, sizeof(int));
356 start_address += sizeof(int);
357
358 /* Write details for each parameter in turn. */
359 paramid = -1;
360 while ((paramid = bms_next_member(params, paramid)) >= 0)
361 {
362 Oid typeOid;
363 int16 typLen;
364 bool typByVal;
365
366 prm = &(estate->es_param_exec_vals[paramid]);
367 typeOid = list_nth_oid(estate->es_plannedstmt->paramExecTypes,
368 paramid);
369
370 /* Write paramid. */
371 memcpy(start_address, ¶mid, sizeof(int));
372 start_address += sizeof(int);
373
374 /* Write datum/isnull */
375 if (OidIsValid(typeOid))
376 get_typlenbyval(typeOid, &typLen, &typByVal);
377 else
378 {
379 /* If no type OID, assume by-value, like copyParamList does. */
380 typLen = sizeof(Datum);
381 typByVal = true;
382 }
383 datumSerialize(prm->value, prm->isnull, typByVal, typLen,
384 &start_address);
385 }
386
387 return handle;
388 }
389
390 /*
391 * Restore specified PARAM_EXEC parameters.
392 */
393 static void
RestoreParamExecParams(char * start_address,EState * estate)394 RestoreParamExecParams(char *start_address, EState *estate)
395 {
396 int nparams;
397 int i;
398 int paramid;
399
400 memcpy(&nparams, start_address, sizeof(int));
401 start_address += sizeof(int);
402
403 for (i = 0; i < nparams; i++)
404 {
405 ParamExecData *prm;
406
407 /* Read paramid */
408 memcpy(¶mid, start_address, sizeof(int));
409 start_address += sizeof(int);
410 prm = &(estate->es_param_exec_vals[paramid]);
411
412 /* Read datum/isnull. */
413 prm->value = datumRestore(&start_address, &prm->isnull);
414 prm->execPlan = NULL;
415 }
416 }
417
418 /*
419 * Initialize the dynamic shared memory segment that will be used to control
420 * parallel execution.
421 */
422 static bool
ExecParallelInitializeDSM(PlanState * planstate,ExecParallelInitializeDSMContext * d)423 ExecParallelInitializeDSM(PlanState *planstate,
424 ExecParallelInitializeDSMContext *d)
425 {
426 if (planstate == NULL)
427 return false;
428
429 /* If instrumentation is enabled, initialize slot for this node. */
430 if (d->instrumentation != NULL)
431 d->instrumentation->plan_node_id[d->nnodes] =
432 planstate->plan->plan_node_id;
433
434 /* Count this node. */
435 d->nnodes++;
436
437 /*
438 * Call initializers for DSM-using plan nodes.
439 *
440 * Most plan nodes won't do anything here, but plan nodes that allocated
441 * DSM may need to initialize shared state in the DSM before parallel
442 * workers are launched. They can allocate the space they previously
443 * estimated using shm_toc_allocate, and add the keys they previously
444 * estimated using shm_toc_insert, in each case targeting pcxt->toc.
445 */
446 switch (nodeTag(planstate))
447 {
448 case T_SeqScanState:
449 if (planstate->plan->parallel_aware)
450 ExecSeqScanInitializeDSM((SeqScanState *) planstate,
451 d->pcxt);
452 break;
453 case T_IndexScanState:
454 if (planstate->plan->parallel_aware)
455 ExecIndexScanInitializeDSM((IndexScanState *) planstate,
456 d->pcxt);
457 break;
458 case T_IndexOnlyScanState:
459 if (planstate->plan->parallel_aware)
460 ExecIndexOnlyScanInitializeDSM((IndexOnlyScanState *) planstate,
461 d->pcxt);
462 break;
463 case T_ForeignScanState:
464 if (planstate->plan->parallel_aware)
465 ExecForeignScanInitializeDSM((ForeignScanState *) planstate,
466 d->pcxt);
467 break;
468 case T_AppendState:
469 if (planstate->plan->parallel_aware)
470 ExecAppendInitializeDSM((AppendState *) planstate,
471 d->pcxt);
472 break;
473 case T_CustomScanState:
474 if (planstate->plan->parallel_aware)
475 ExecCustomScanInitializeDSM((CustomScanState *) planstate,
476 d->pcxt);
477 break;
478 case T_BitmapHeapScanState:
479 if (planstate->plan->parallel_aware)
480 ExecBitmapHeapInitializeDSM((BitmapHeapScanState *) planstate,
481 d->pcxt);
482 break;
483 case T_HashJoinState:
484 if (planstate->plan->parallel_aware)
485 ExecHashJoinInitializeDSM((HashJoinState *) planstate,
486 d->pcxt);
487 break;
488 case T_HashState:
489 /* even when not parallel-aware, for EXPLAIN ANALYZE */
490 ExecHashInitializeDSM((HashState *) planstate, d->pcxt);
491 break;
492 case T_SortState:
493 /* even when not parallel-aware, for EXPLAIN ANALYZE */
494 ExecSortInitializeDSM((SortState *) planstate, d->pcxt);
495 break;
496
497 default:
498 break;
499 }
500
501 return planstate_tree_walker(planstate, ExecParallelInitializeDSM, d);
502 }
503
504 /*
505 * It sets up the response queues for backend workers to return tuples
506 * to the main backend and start the workers.
507 */
508 static shm_mq_handle **
ExecParallelSetupTupleQueues(ParallelContext * pcxt,bool reinitialize)509 ExecParallelSetupTupleQueues(ParallelContext *pcxt, bool reinitialize)
510 {
511 shm_mq_handle **responseq;
512 char *tqueuespace;
513 int i;
514
515 /* Skip this if no workers. */
516 if (pcxt->nworkers == 0)
517 return NULL;
518
519 /* Allocate memory for shared memory queue handles. */
520 responseq = (shm_mq_handle **)
521 palloc(pcxt->nworkers * sizeof(shm_mq_handle *));
522
523 /*
524 * If not reinitializing, allocate space from the DSM for the queues;
525 * otherwise, find the already allocated space.
526 */
527 if (!reinitialize)
528 tqueuespace =
529 shm_toc_allocate(pcxt->toc,
530 mul_size(PARALLEL_TUPLE_QUEUE_SIZE,
531 pcxt->nworkers));
532 else
533 tqueuespace = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_TUPLE_QUEUE, false);
534
535 /* Create the queues, and become the receiver for each. */
536 for (i = 0; i < pcxt->nworkers; ++i)
537 {
538 shm_mq *mq;
539
540 mq = shm_mq_create(tqueuespace +
541 ((Size) i) * PARALLEL_TUPLE_QUEUE_SIZE,
542 (Size) PARALLEL_TUPLE_QUEUE_SIZE);
543
544 shm_mq_set_receiver(mq, MyProc);
545 responseq[i] = shm_mq_attach(mq, pcxt->seg, NULL);
546 }
547
548 /* Add array of queues to shm_toc, so others can find it. */
549 if (!reinitialize)
550 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TUPLE_QUEUE, tqueuespace);
551
552 /* Return array of handles. */
553 return responseq;
554 }
555
556 /*
557 * Sets up the required infrastructure for backend workers to perform
558 * execution and return results to the main backend.
559 */
560 ParallelExecutorInfo *
ExecInitParallelPlan(PlanState * planstate,EState * estate,Bitmapset * sendParams,int nworkers,int64 tuples_needed)561 ExecInitParallelPlan(PlanState *planstate, EState *estate,
562 Bitmapset *sendParams, int nworkers,
563 int64 tuples_needed)
564 {
565 ParallelExecutorInfo *pei;
566 ParallelContext *pcxt;
567 ExecParallelEstimateContext e;
568 ExecParallelInitializeDSMContext d;
569 FixedParallelExecutorState *fpes;
570 char *pstmt_data;
571 char *pstmt_space;
572 char *paramlistinfo_space;
573 BufferUsage *bufusage_space;
574 SharedExecutorInstrumentation *instrumentation = NULL;
575 SharedJitInstrumentation *jit_instrumentation = NULL;
576 int pstmt_len;
577 int paramlistinfo_len;
578 int instrumentation_len = 0;
579 int jit_instrumentation_len = 0;
580 int instrument_offset = 0;
581 Size dsa_minsize = dsa_minimum_size();
582 char *query_string;
583 int query_len;
584
585 /*
586 * Force any initplan outputs that we're going to pass to workers to be
587 * evaluated, if they weren't already.
588 *
589 * For simplicity, we use the EState's per-output-tuple ExprContext here.
590 * That risks intra-query memory leakage, since we might pass through here
591 * many times before that ExprContext gets reset; but ExecSetParamPlan
592 * doesn't normally leak any memory in the context (see its comments), so
593 * it doesn't seem worth complicating this function's API to pass it a
594 * shorter-lived ExprContext. This might need to change someday.
595 */
596 ExecSetParamPlanMulti(sendParams, GetPerTupleExprContext(estate));
597
598 /* Allocate object for return value. */
599 pei = palloc0(sizeof(ParallelExecutorInfo));
600 pei->finished = false;
601 pei->planstate = planstate;
602
603 /* Fix up and serialize plan to be sent to workers. */
604 pstmt_data = ExecSerializePlan(planstate->plan, estate);
605
606 /* Create a parallel context. */
607 pcxt = CreateParallelContext("postgres", "ParallelQueryMain", nworkers);
608 pei->pcxt = pcxt;
609
610 /*
611 * Before telling the parallel context to create a dynamic shared memory
612 * segment, we need to figure out how big it should be. Estimate space
613 * for the various things we need to store.
614 */
615
616 /* Estimate space for fixed-size state. */
617 shm_toc_estimate_chunk(&pcxt->estimator,
618 sizeof(FixedParallelExecutorState));
619 shm_toc_estimate_keys(&pcxt->estimator, 1);
620
621 /* Estimate space for query text. */
622 query_len = strlen(estate->es_sourceText);
623 shm_toc_estimate_chunk(&pcxt->estimator, query_len + 1);
624 shm_toc_estimate_keys(&pcxt->estimator, 1);
625
626 /* Estimate space for serialized PlannedStmt. */
627 pstmt_len = strlen(pstmt_data) + 1;
628 shm_toc_estimate_chunk(&pcxt->estimator, pstmt_len);
629 shm_toc_estimate_keys(&pcxt->estimator, 1);
630
631 /* Estimate space for serialized ParamListInfo. */
632 paramlistinfo_len = EstimateParamListSpace(estate->es_param_list_info);
633 shm_toc_estimate_chunk(&pcxt->estimator, paramlistinfo_len);
634 shm_toc_estimate_keys(&pcxt->estimator, 1);
635
636 /*
637 * Estimate space for BufferUsage.
638 *
639 * If EXPLAIN is not in use and there are no extensions loaded that care,
640 * we could skip this. But we have no way of knowing whether anyone's
641 * looking at pgBufferUsage, so do it unconditionally.
642 */
643 shm_toc_estimate_chunk(&pcxt->estimator,
644 mul_size(sizeof(BufferUsage), pcxt->nworkers));
645 shm_toc_estimate_keys(&pcxt->estimator, 1);
646
647 /* Estimate space for tuple queues. */
648 shm_toc_estimate_chunk(&pcxt->estimator,
649 mul_size(PARALLEL_TUPLE_QUEUE_SIZE, pcxt->nworkers));
650 shm_toc_estimate_keys(&pcxt->estimator, 1);
651
652 /*
653 * Give parallel-aware nodes a chance to add to the estimates, and get a
654 * count of how many PlanState nodes there are.
655 */
656 e.pcxt = pcxt;
657 e.nnodes = 0;
658 ExecParallelEstimate(planstate, &e);
659
660 /* Estimate space for instrumentation, if required. */
661 if (estate->es_instrument)
662 {
663 instrumentation_len =
664 offsetof(SharedExecutorInstrumentation, plan_node_id) +
665 sizeof(int) * e.nnodes;
666 instrumentation_len = MAXALIGN(instrumentation_len);
667 instrument_offset = instrumentation_len;
668 instrumentation_len +=
669 mul_size(sizeof(Instrumentation),
670 mul_size(e.nnodes, nworkers));
671 shm_toc_estimate_chunk(&pcxt->estimator, instrumentation_len);
672 shm_toc_estimate_keys(&pcxt->estimator, 1);
673
674 /* Estimate space for JIT instrumentation, if required. */
675 if (estate->es_jit_flags != PGJIT_NONE)
676 {
677 jit_instrumentation_len =
678 offsetof(SharedJitInstrumentation, jit_instr) +
679 sizeof(JitInstrumentation) * nworkers;
680 shm_toc_estimate_chunk(&pcxt->estimator, jit_instrumentation_len);
681 shm_toc_estimate_keys(&pcxt->estimator, 1);
682 }
683 }
684
685 /* Estimate space for DSA area. */
686 shm_toc_estimate_chunk(&pcxt->estimator, dsa_minsize);
687 shm_toc_estimate_keys(&pcxt->estimator, 1);
688
689 /* Everyone's had a chance to ask for space, so now create the DSM. */
690 InitializeParallelDSM(pcxt);
691
692 /*
693 * OK, now we have a dynamic shared memory segment, and it should be big
694 * enough to store all of the data we estimated we would want to put into
695 * it, plus whatever general stuff (not specifically executor-related) the
696 * ParallelContext itself needs to store there. None of the space we
697 * asked for has been allocated or initialized yet, though, so do that.
698 */
699
700 /* Store fixed-size state. */
701 fpes = shm_toc_allocate(pcxt->toc, sizeof(FixedParallelExecutorState));
702 fpes->tuples_needed = tuples_needed;
703 fpes->param_exec = InvalidDsaPointer;
704 fpes->eflags = estate->es_top_eflags;
705 fpes->jit_flags = estate->es_jit_flags;
706 shm_toc_insert(pcxt->toc, PARALLEL_KEY_EXECUTOR_FIXED, fpes);
707
708 /* Store query string */
709 query_string = shm_toc_allocate(pcxt->toc, query_len + 1);
710 memcpy(query_string, estate->es_sourceText, query_len + 1);
711 shm_toc_insert(pcxt->toc, PARALLEL_KEY_QUERY_TEXT, query_string);
712
713 /* Store serialized PlannedStmt. */
714 pstmt_space = shm_toc_allocate(pcxt->toc, pstmt_len);
715 memcpy(pstmt_space, pstmt_data, pstmt_len);
716 shm_toc_insert(pcxt->toc, PARALLEL_KEY_PLANNEDSTMT, pstmt_space);
717
718 /* Store serialized ParamListInfo. */
719 paramlistinfo_space = shm_toc_allocate(pcxt->toc, paramlistinfo_len);
720 shm_toc_insert(pcxt->toc, PARALLEL_KEY_PARAMLISTINFO, paramlistinfo_space);
721 SerializeParamList(estate->es_param_list_info, ¶mlistinfo_space);
722
723 /* Allocate space for each worker's BufferUsage; no need to initialize. */
724 bufusage_space = shm_toc_allocate(pcxt->toc,
725 mul_size(sizeof(BufferUsage), pcxt->nworkers));
726 shm_toc_insert(pcxt->toc, PARALLEL_KEY_BUFFER_USAGE, bufusage_space);
727 pei->buffer_usage = bufusage_space;
728
729 /* Set up the tuple queues that the workers will write into. */
730 pei->tqueue = ExecParallelSetupTupleQueues(pcxt, false);
731
732 /* We don't need the TupleQueueReaders yet, though. */
733 pei->reader = NULL;
734
735 /*
736 * If instrumentation options were supplied, allocate space for the data.
737 * It only gets partially initialized here; the rest happens during
738 * ExecParallelInitializeDSM.
739 */
740 if (estate->es_instrument)
741 {
742 Instrumentation *instrument;
743 int i;
744
745 instrumentation = shm_toc_allocate(pcxt->toc, instrumentation_len);
746 instrumentation->instrument_options = estate->es_instrument;
747 instrumentation->instrument_offset = instrument_offset;
748 instrumentation->num_workers = nworkers;
749 instrumentation->num_plan_nodes = e.nnodes;
750 instrument = GetInstrumentationArray(instrumentation);
751 for (i = 0; i < nworkers * e.nnodes; ++i)
752 InstrInit(&instrument[i], estate->es_instrument);
753 shm_toc_insert(pcxt->toc, PARALLEL_KEY_INSTRUMENTATION,
754 instrumentation);
755 pei->instrumentation = instrumentation;
756
757 if (estate->es_jit_flags != PGJIT_NONE)
758 {
759 jit_instrumentation = shm_toc_allocate(pcxt->toc,
760 jit_instrumentation_len);
761 jit_instrumentation->num_workers = nworkers;
762 memset(jit_instrumentation->jit_instr, 0,
763 sizeof(JitInstrumentation) * nworkers);
764 shm_toc_insert(pcxt->toc, PARALLEL_KEY_JIT_INSTRUMENTATION,
765 jit_instrumentation);
766 pei->jit_instrumentation = jit_instrumentation;
767 }
768 }
769
770 /*
771 * Create a DSA area that can be used by the leader and all workers.
772 * (However, if we failed to create a DSM and are using private memory
773 * instead, then skip this.)
774 */
775 if (pcxt->seg != NULL)
776 {
777 char *area_space;
778
779 area_space = shm_toc_allocate(pcxt->toc, dsa_minsize);
780 shm_toc_insert(pcxt->toc, PARALLEL_KEY_DSA, area_space);
781 pei->area = dsa_create_in_place(area_space, dsa_minsize,
782 LWTRANCHE_PARALLEL_QUERY_DSA,
783 pcxt->seg);
784
785 /*
786 * Serialize parameters, if any, using DSA storage. We don't dare use
787 * the main parallel query DSM for this because we might relaunch
788 * workers after the values have changed (and thus the amount of
789 * storage required has changed).
790 */
791 if (!bms_is_empty(sendParams))
792 {
793 pei->param_exec = SerializeParamExecParams(estate, sendParams,
794 pei->area);
795 fpes->param_exec = pei->param_exec;
796 }
797 }
798
799 /*
800 * Give parallel-aware nodes a chance to initialize their shared data.
801 * This also initializes the elements of instrumentation->ps_instrument,
802 * if it exists.
803 */
804 d.pcxt = pcxt;
805 d.instrumentation = instrumentation;
806 d.nnodes = 0;
807
808 /* Install our DSA area while initializing the plan. */
809 estate->es_query_dsa = pei->area;
810 ExecParallelInitializeDSM(planstate, &d);
811 estate->es_query_dsa = NULL;
812
813 /*
814 * Make sure that the world hasn't shifted under our feet. This could
815 * probably just be an Assert(), but let's be conservative for now.
816 */
817 if (e.nnodes != d.nnodes)
818 elog(ERROR, "inconsistent count of PlanState nodes");
819
820 /* OK, we're ready to rock and roll. */
821 return pei;
822 }
823
824 /*
825 * Set up tuple queue readers to read the results of a parallel subplan.
826 *
827 * This is separate from ExecInitParallelPlan() because we can launch the
828 * worker processes and let them start doing something before we do this.
829 */
830 void
ExecParallelCreateReaders(ParallelExecutorInfo * pei)831 ExecParallelCreateReaders(ParallelExecutorInfo *pei)
832 {
833 int nworkers = pei->pcxt->nworkers_launched;
834 int i;
835
836 Assert(pei->reader == NULL);
837
838 if (nworkers > 0)
839 {
840 pei->reader = (TupleQueueReader **)
841 palloc(nworkers * sizeof(TupleQueueReader *));
842
843 for (i = 0; i < nworkers; i++)
844 {
845 shm_mq_set_handle(pei->tqueue[i],
846 pei->pcxt->worker[i].bgwhandle);
847 pei->reader[i] = CreateTupleQueueReader(pei->tqueue[i]);
848 }
849 }
850 }
851
852 /*
853 * Re-initialize the parallel executor shared memory state before launching
854 * a fresh batch of workers.
855 */
856 void
ExecParallelReinitialize(PlanState * planstate,ParallelExecutorInfo * pei,Bitmapset * sendParams)857 ExecParallelReinitialize(PlanState *planstate,
858 ParallelExecutorInfo *pei,
859 Bitmapset *sendParams)
860 {
861 EState *estate = planstate->state;
862 FixedParallelExecutorState *fpes;
863
864 /* Old workers must already be shut down */
865 Assert(pei->finished);
866
867 /*
868 * Force any initplan outputs that we're going to pass to workers to be
869 * evaluated, if they weren't already (see comments in
870 * ExecInitParallelPlan).
871 */
872 ExecSetParamPlanMulti(sendParams, GetPerTupleExprContext(estate));
873
874 ReinitializeParallelDSM(pei->pcxt);
875 pei->tqueue = ExecParallelSetupTupleQueues(pei->pcxt, true);
876 pei->reader = NULL;
877 pei->finished = false;
878
879 fpes = shm_toc_lookup(pei->pcxt->toc, PARALLEL_KEY_EXECUTOR_FIXED, false);
880
881 /* Free any serialized parameters from the last round. */
882 if (DsaPointerIsValid(fpes->param_exec))
883 {
884 dsa_free(pei->area, fpes->param_exec);
885 fpes->param_exec = InvalidDsaPointer;
886 }
887
888 /* Serialize current parameter values if required. */
889 if (!bms_is_empty(sendParams))
890 {
891 pei->param_exec = SerializeParamExecParams(estate, sendParams,
892 pei->area);
893 fpes->param_exec = pei->param_exec;
894 }
895
896 /* Traverse plan tree and let each child node reset associated state. */
897 estate->es_query_dsa = pei->area;
898 ExecParallelReInitializeDSM(planstate, pei->pcxt);
899 estate->es_query_dsa = NULL;
900 }
901
902 /*
903 * Traverse plan tree to reinitialize per-node dynamic shared memory state
904 */
905 static bool
ExecParallelReInitializeDSM(PlanState * planstate,ParallelContext * pcxt)906 ExecParallelReInitializeDSM(PlanState *planstate,
907 ParallelContext *pcxt)
908 {
909 if (planstate == NULL)
910 return false;
911
912 /*
913 * Call reinitializers for DSM-using plan nodes.
914 */
915 switch (nodeTag(planstate))
916 {
917 case T_SeqScanState:
918 if (planstate->plan->parallel_aware)
919 ExecSeqScanReInitializeDSM((SeqScanState *) planstate,
920 pcxt);
921 break;
922 case T_IndexScanState:
923 if (planstate->plan->parallel_aware)
924 ExecIndexScanReInitializeDSM((IndexScanState *) planstate,
925 pcxt);
926 break;
927 case T_IndexOnlyScanState:
928 if (planstate->plan->parallel_aware)
929 ExecIndexOnlyScanReInitializeDSM((IndexOnlyScanState *) planstate,
930 pcxt);
931 break;
932 case T_ForeignScanState:
933 if (planstate->plan->parallel_aware)
934 ExecForeignScanReInitializeDSM((ForeignScanState *) planstate,
935 pcxt);
936 break;
937 case T_AppendState:
938 if (planstate->plan->parallel_aware)
939 ExecAppendReInitializeDSM((AppendState *) planstate, pcxt);
940 break;
941 case T_CustomScanState:
942 if (planstate->plan->parallel_aware)
943 ExecCustomScanReInitializeDSM((CustomScanState *) planstate,
944 pcxt);
945 break;
946 case T_BitmapHeapScanState:
947 if (planstate->plan->parallel_aware)
948 ExecBitmapHeapReInitializeDSM((BitmapHeapScanState *) planstate,
949 pcxt);
950 break;
951 case T_HashJoinState:
952 if (planstate->plan->parallel_aware)
953 ExecHashJoinReInitializeDSM((HashJoinState *) planstate,
954 pcxt);
955 break;
956 case T_HashState:
957 case T_SortState:
958 /* these nodes have DSM state, but no reinitialization is required */
959 break;
960
961 default:
962 break;
963 }
964
965 return planstate_tree_walker(planstate, ExecParallelReInitializeDSM, pcxt);
966 }
967
968 /*
969 * Copy instrumentation information about this node and its descendants from
970 * dynamic shared memory.
971 */
972 static bool
ExecParallelRetrieveInstrumentation(PlanState * planstate,SharedExecutorInstrumentation * instrumentation)973 ExecParallelRetrieveInstrumentation(PlanState *planstate,
974 SharedExecutorInstrumentation *instrumentation)
975 {
976 Instrumentation *instrument;
977 int i;
978 int n;
979 int ibytes;
980 int plan_node_id = planstate->plan->plan_node_id;
981 MemoryContext oldcontext;
982
983 /* Find the instrumentation for this node. */
984 for (i = 0; i < instrumentation->num_plan_nodes; ++i)
985 if (instrumentation->plan_node_id[i] == plan_node_id)
986 break;
987 if (i >= instrumentation->num_plan_nodes)
988 elog(ERROR, "plan node %d not found", plan_node_id);
989
990 /* Accumulate the statistics from all workers. */
991 instrument = GetInstrumentationArray(instrumentation);
992 instrument += i * instrumentation->num_workers;
993 for (n = 0; n < instrumentation->num_workers; ++n)
994 InstrAggNode(planstate->instrument, &instrument[n]);
995
996 /*
997 * Also store the per-worker detail.
998 *
999 * Worker instrumentation should be allocated in the same context as the
1000 * regular instrumentation information, which is the per-query context.
1001 * Switch into per-query memory context.
1002 */
1003 oldcontext = MemoryContextSwitchTo(planstate->state->es_query_cxt);
1004 ibytes = mul_size(instrumentation->num_workers, sizeof(Instrumentation));
1005 planstate->worker_instrument =
1006 palloc(ibytes + offsetof(WorkerInstrumentation, instrument));
1007 MemoryContextSwitchTo(oldcontext);
1008
1009 planstate->worker_instrument->num_workers = instrumentation->num_workers;
1010 memcpy(&planstate->worker_instrument->instrument, instrument, ibytes);
1011
1012 /* Perform any node-type-specific work that needs to be done. */
1013 switch (nodeTag(planstate))
1014 {
1015 case T_SortState:
1016 ExecSortRetrieveInstrumentation((SortState *) planstate);
1017 break;
1018 case T_HashState:
1019 ExecHashRetrieveInstrumentation((HashState *) planstate);
1020 break;
1021 default:
1022 break;
1023 }
1024
1025 return planstate_tree_walker(planstate, ExecParallelRetrieveInstrumentation,
1026 instrumentation);
1027 }
1028
1029 /*
1030 * Add up the workers' JIT instrumentation from dynamic shared memory.
1031 */
1032 static void
ExecParallelRetrieveJitInstrumentation(PlanState * planstate,SharedJitInstrumentation * shared_jit)1033 ExecParallelRetrieveJitInstrumentation(PlanState *planstate,
1034 SharedJitInstrumentation *shared_jit)
1035 {
1036 JitInstrumentation *combined;
1037 int ibytes;
1038
1039 int n;
1040
1041 /*
1042 * Accumulate worker JIT instrumentation into the combined JIT
1043 * instrumentation, allocating it if required.
1044 */
1045 if (!planstate->state->es_jit_worker_instr)
1046 planstate->state->es_jit_worker_instr =
1047 MemoryContextAllocZero(planstate->state->es_query_cxt, sizeof(JitInstrumentation));
1048 combined = planstate->state->es_jit_worker_instr;
1049
1050 /* Accumulate all the workers' instrumentations. */
1051 for (n = 0; n < shared_jit->num_workers; ++n)
1052 InstrJitAgg(combined, &shared_jit->jit_instr[n]);
1053
1054 /*
1055 * Store the per-worker detail.
1056 *
1057 * Similar to ExecParallelRetrieveInstrumentation(), allocate the
1058 * instrumentation in per-query context.
1059 */
1060 ibytes = offsetof(SharedJitInstrumentation, jit_instr)
1061 + mul_size(shared_jit->num_workers, sizeof(JitInstrumentation));
1062 planstate->worker_jit_instrument =
1063 MemoryContextAlloc(planstate->state->es_query_cxt, ibytes);
1064
1065 memcpy(planstate->worker_jit_instrument, shared_jit, ibytes);
1066 }
1067
1068 /*
1069 * Finish parallel execution. We wait for parallel workers to finish, and
1070 * accumulate their buffer usage.
1071 */
1072 void
ExecParallelFinish(ParallelExecutorInfo * pei)1073 ExecParallelFinish(ParallelExecutorInfo *pei)
1074 {
1075 int nworkers = pei->pcxt->nworkers_launched;
1076 int i;
1077
1078 /* Make this be a no-op if called twice in a row. */
1079 if (pei->finished)
1080 return;
1081
1082 /*
1083 * Detach from tuple queues ASAP, so that any still-active workers will
1084 * notice that no further results are wanted.
1085 */
1086 if (pei->tqueue != NULL)
1087 {
1088 for (i = 0; i < nworkers; i++)
1089 shm_mq_detach(pei->tqueue[i]);
1090 pfree(pei->tqueue);
1091 pei->tqueue = NULL;
1092 }
1093
1094 /*
1095 * While we're waiting for the workers to finish, let's get rid of the
1096 * tuple queue readers. (Any other local cleanup could be done here too.)
1097 */
1098 if (pei->reader != NULL)
1099 {
1100 for (i = 0; i < nworkers; i++)
1101 DestroyTupleQueueReader(pei->reader[i]);
1102 pfree(pei->reader);
1103 pei->reader = NULL;
1104 }
1105
1106 /* Now wait for the workers to finish. */
1107 WaitForParallelWorkersToFinish(pei->pcxt);
1108
1109 /*
1110 * Next, accumulate buffer usage. (This must wait for the workers to
1111 * finish, or we might get incomplete data.)
1112 */
1113 for (i = 0; i < nworkers; i++)
1114 InstrAccumParallelQuery(&pei->buffer_usage[i]);
1115
1116 pei->finished = true;
1117 }
1118
1119 /*
1120 * Accumulate instrumentation, and then clean up whatever ParallelExecutorInfo
1121 * resources still exist after ExecParallelFinish. We separate these
1122 * routines because someone might want to examine the contents of the DSM
1123 * after ExecParallelFinish and before calling this routine.
1124 */
1125 void
ExecParallelCleanup(ParallelExecutorInfo * pei)1126 ExecParallelCleanup(ParallelExecutorInfo *pei)
1127 {
1128 /* Accumulate instrumentation, if any. */
1129 if (pei->instrumentation)
1130 ExecParallelRetrieveInstrumentation(pei->planstate,
1131 pei->instrumentation);
1132
1133 /* Accumulate JIT instrumentation, if any. */
1134 if (pei->jit_instrumentation)
1135 ExecParallelRetrieveJitInstrumentation(pei->planstate,
1136 pei->jit_instrumentation);
1137
1138 /* Free any serialized parameters. */
1139 if (DsaPointerIsValid(pei->param_exec))
1140 {
1141 dsa_free(pei->area, pei->param_exec);
1142 pei->param_exec = InvalidDsaPointer;
1143 }
1144 if (pei->area != NULL)
1145 {
1146 dsa_detach(pei->area);
1147 pei->area = NULL;
1148 }
1149 if (pei->pcxt != NULL)
1150 {
1151 DestroyParallelContext(pei->pcxt);
1152 pei->pcxt = NULL;
1153 }
1154 pfree(pei);
1155 }
1156
1157 /*
1158 * Create a DestReceiver to write tuples we produce to the shm_mq designated
1159 * for that purpose.
1160 */
1161 static DestReceiver *
ExecParallelGetReceiver(dsm_segment * seg,shm_toc * toc)1162 ExecParallelGetReceiver(dsm_segment *seg, shm_toc *toc)
1163 {
1164 char *mqspace;
1165 shm_mq *mq;
1166
1167 mqspace = shm_toc_lookup(toc, PARALLEL_KEY_TUPLE_QUEUE, false);
1168 mqspace += ParallelWorkerNumber * PARALLEL_TUPLE_QUEUE_SIZE;
1169 mq = (shm_mq *) mqspace;
1170 shm_mq_set_sender(mq, MyProc);
1171 return CreateTupleQueueDestReceiver(shm_mq_attach(mq, seg, NULL));
1172 }
1173
1174 /*
1175 * Create a QueryDesc for the PlannedStmt we are to execute, and return it.
1176 */
1177 static QueryDesc *
ExecParallelGetQueryDesc(shm_toc * toc,DestReceiver * receiver,int instrument_options)1178 ExecParallelGetQueryDesc(shm_toc *toc, DestReceiver *receiver,
1179 int instrument_options)
1180 {
1181 char *pstmtspace;
1182 char *paramspace;
1183 PlannedStmt *pstmt;
1184 ParamListInfo paramLI;
1185 char *queryString;
1186
1187 /* Get the query string from shared memory */
1188 queryString = shm_toc_lookup(toc, PARALLEL_KEY_QUERY_TEXT, false);
1189
1190 /* Reconstruct leader-supplied PlannedStmt. */
1191 pstmtspace = shm_toc_lookup(toc, PARALLEL_KEY_PLANNEDSTMT, false);
1192 pstmt = (PlannedStmt *) stringToNode(pstmtspace);
1193
1194 /* Reconstruct ParamListInfo. */
1195 paramspace = shm_toc_lookup(toc, PARALLEL_KEY_PARAMLISTINFO, false);
1196 paramLI = RestoreParamList(¶mspace);
1197
1198 /* Create a QueryDesc for the query. */
1199 return CreateQueryDesc(pstmt,
1200 queryString,
1201 GetActiveSnapshot(), InvalidSnapshot,
1202 receiver, paramLI, NULL, instrument_options);
1203 }
1204
1205 /*
1206 * Copy instrumentation information from this node and its descendants into
1207 * dynamic shared memory, so that the parallel leader can retrieve it.
1208 */
1209 static bool
ExecParallelReportInstrumentation(PlanState * planstate,SharedExecutorInstrumentation * instrumentation)1210 ExecParallelReportInstrumentation(PlanState *planstate,
1211 SharedExecutorInstrumentation *instrumentation)
1212 {
1213 int i;
1214 int plan_node_id = planstate->plan->plan_node_id;
1215 Instrumentation *instrument;
1216
1217 InstrEndLoop(planstate->instrument);
1218
1219 /*
1220 * If we shuffled the plan_node_id values in ps_instrument into sorted
1221 * order, we could use binary search here. This might matter someday if
1222 * we're pushing down sufficiently large plan trees. For now, do it the
1223 * slow, dumb way.
1224 */
1225 for (i = 0; i < instrumentation->num_plan_nodes; ++i)
1226 if (instrumentation->plan_node_id[i] == plan_node_id)
1227 break;
1228 if (i >= instrumentation->num_plan_nodes)
1229 elog(ERROR, "plan node %d not found", plan_node_id);
1230
1231 /*
1232 * Add our statistics to the per-node, per-worker totals. It's possible
1233 * that this could happen more than once if we relaunched workers.
1234 */
1235 instrument = GetInstrumentationArray(instrumentation);
1236 instrument += i * instrumentation->num_workers;
1237 Assert(IsParallelWorker());
1238 Assert(ParallelWorkerNumber < instrumentation->num_workers);
1239 InstrAggNode(&instrument[ParallelWorkerNumber], planstate->instrument);
1240
1241 return planstate_tree_walker(planstate, ExecParallelReportInstrumentation,
1242 instrumentation);
1243 }
1244
1245 /*
1246 * Initialize the PlanState and its descendants with the information
1247 * retrieved from shared memory. This has to be done once the PlanState
1248 * is allocated and initialized by executor; that is, after ExecutorStart().
1249 */
1250 static bool
ExecParallelInitializeWorker(PlanState * planstate,ParallelWorkerContext * pwcxt)1251 ExecParallelInitializeWorker(PlanState *planstate, ParallelWorkerContext *pwcxt)
1252 {
1253 if (planstate == NULL)
1254 return false;
1255
1256 switch (nodeTag(planstate))
1257 {
1258 case T_SeqScanState:
1259 if (planstate->plan->parallel_aware)
1260 ExecSeqScanInitializeWorker((SeqScanState *) planstate, pwcxt);
1261 break;
1262 case T_IndexScanState:
1263 if (planstate->plan->parallel_aware)
1264 ExecIndexScanInitializeWorker((IndexScanState *) planstate,
1265 pwcxt);
1266 break;
1267 case T_IndexOnlyScanState:
1268 if (planstate->plan->parallel_aware)
1269 ExecIndexOnlyScanInitializeWorker((IndexOnlyScanState *) planstate,
1270 pwcxt);
1271 break;
1272 case T_ForeignScanState:
1273 if (planstate->plan->parallel_aware)
1274 ExecForeignScanInitializeWorker((ForeignScanState *) planstate,
1275 pwcxt);
1276 break;
1277 case T_AppendState:
1278 if (planstate->plan->parallel_aware)
1279 ExecAppendInitializeWorker((AppendState *) planstate, pwcxt);
1280 break;
1281 case T_CustomScanState:
1282 if (planstate->plan->parallel_aware)
1283 ExecCustomScanInitializeWorker((CustomScanState *) planstate,
1284 pwcxt);
1285 break;
1286 case T_BitmapHeapScanState:
1287 if (planstate->plan->parallel_aware)
1288 ExecBitmapHeapInitializeWorker((BitmapHeapScanState *) planstate,
1289 pwcxt);
1290 break;
1291 case T_HashJoinState:
1292 if (planstate->plan->parallel_aware)
1293 ExecHashJoinInitializeWorker((HashJoinState *) planstate,
1294 pwcxt);
1295 break;
1296 case T_HashState:
1297 /* even when not parallel-aware, for EXPLAIN ANALYZE */
1298 ExecHashInitializeWorker((HashState *) planstate, pwcxt);
1299 break;
1300 case T_SortState:
1301 /* even when not parallel-aware, for EXPLAIN ANALYZE */
1302 ExecSortInitializeWorker((SortState *) planstate, pwcxt);
1303 break;
1304
1305 default:
1306 break;
1307 }
1308
1309 return planstate_tree_walker(planstate, ExecParallelInitializeWorker,
1310 pwcxt);
1311 }
1312
1313 /*
1314 * Main entrypoint for parallel query worker processes.
1315 *
1316 * We reach this function from ParallelWorkerMain, so the setup necessary to
1317 * create a sensible parallel environment has already been done;
1318 * ParallelWorkerMain worries about stuff like the transaction state, combo
1319 * CID mappings, and GUC values, so we don't need to deal with any of that
1320 * here.
1321 *
1322 * Our job is to deal with concerns specific to the executor. The parallel
1323 * group leader will have stored a serialized PlannedStmt, and it's our job
1324 * to execute that plan and write the resulting tuples to the appropriate
1325 * tuple queue. Various bits of supporting information that we need in order
1326 * to do this are also stored in the dsm_segment and can be accessed through
1327 * the shm_toc.
1328 */
1329 void
ParallelQueryMain(dsm_segment * seg,shm_toc * toc)1330 ParallelQueryMain(dsm_segment *seg, shm_toc *toc)
1331 {
1332 FixedParallelExecutorState *fpes;
1333 BufferUsage *buffer_usage;
1334 DestReceiver *receiver;
1335 QueryDesc *queryDesc;
1336 SharedExecutorInstrumentation *instrumentation;
1337 SharedJitInstrumentation *jit_instrumentation;
1338 int instrument_options = 0;
1339 void *area_space;
1340 dsa_area *area;
1341 ParallelWorkerContext pwcxt;
1342
1343 /* Get fixed-size state. */
1344 fpes = shm_toc_lookup(toc, PARALLEL_KEY_EXECUTOR_FIXED, false);
1345
1346 /* Set up DestReceiver, SharedExecutorInstrumentation, and QueryDesc. */
1347 receiver = ExecParallelGetReceiver(seg, toc);
1348 instrumentation = shm_toc_lookup(toc, PARALLEL_KEY_INSTRUMENTATION, true);
1349 if (instrumentation != NULL)
1350 instrument_options = instrumentation->instrument_options;
1351 jit_instrumentation = shm_toc_lookup(toc, PARALLEL_KEY_JIT_INSTRUMENTATION,
1352 true);
1353 queryDesc = ExecParallelGetQueryDesc(toc, receiver, instrument_options);
1354
1355 /* Setting debug_query_string for individual workers */
1356 debug_query_string = queryDesc->sourceText;
1357
1358 /* Report workers' query for monitoring purposes */
1359 pgstat_report_activity(STATE_RUNNING, debug_query_string);
1360
1361 /* Attach to the dynamic shared memory area. */
1362 area_space = shm_toc_lookup(toc, PARALLEL_KEY_DSA, false);
1363 area = dsa_attach_in_place(area_space, seg);
1364
1365 /* Start up the executor */
1366 queryDesc->plannedstmt->jitFlags = fpes->jit_flags;
1367 ExecutorStart(queryDesc, fpes->eflags);
1368
1369 /* Special executor initialization steps for parallel workers */
1370 queryDesc->planstate->state->es_query_dsa = area;
1371 if (DsaPointerIsValid(fpes->param_exec))
1372 {
1373 char *paramexec_space;
1374
1375 paramexec_space = dsa_get_address(area, fpes->param_exec);
1376 RestoreParamExecParams(paramexec_space, queryDesc->estate);
1377
1378 }
1379 pwcxt.toc = toc;
1380 pwcxt.seg = seg;
1381 ExecParallelInitializeWorker(queryDesc->planstate, &pwcxt);
1382
1383 /* Pass down any tuple bound */
1384 ExecSetTupleBound(fpes->tuples_needed, queryDesc->planstate);
1385
1386 /*
1387 * Prepare to track buffer usage during query execution.
1388 *
1389 * We do this after starting up the executor to match what happens in the
1390 * leader, which also doesn't count buffer accesses that occur during
1391 * executor startup.
1392 */
1393 InstrStartParallelQuery();
1394
1395 /*
1396 * Run the plan. If we specified a tuple bound, be careful not to demand
1397 * more tuples than that.
1398 */
1399 ExecutorRun(queryDesc,
1400 ForwardScanDirection,
1401 fpes->tuples_needed < 0 ? (int64) 0 : fpes->tuples_needed,
1402 true);
1403
1404 /* Shut down the executor */
1405 ExecutorFinish(queryDesc);
1406
1407 /* Report buffer usage during parallel execution. */
1408 buffer_usage = shm_toc_lookup(toc, PARALLEL_KEY_BUFFER_USAGE, false);
1409 InstrEndParallelQuery(&buffer_usage[ParallelWorkerNumber]);
1410
1411 /* Report instrumentation data if any instrumentation options are set. */
1412 if (instrumentation != NULL)
1413 ExecParallelReportInstrumentation(queryDesc->planstate,
1414 instrumentation);
1415
1416 /* Report JIT instrumentation data if any */
1417 if (queryDesc->estate->es_jit && jit_instrumentation != NULL)
1418 {
1419 Assert(ParallelWorkerNumber < jit_instrumentation->num_workers);
1420 jit_instrumentation->jit_instr[ParallelWorkerNumber] =
1421 queryDesc->estate->es_jit->instr;
1422 }
1423
1424 /* Must do this after capturing instrumentation. */
1425 ExecutorEnd(queryDesc);
1426
1427 /* Cleanup. */
1428 dsa_detach(area);
1429 FreeQueryDesc(queryDesc);
1430 receiver->rDestroy(receiver);
1431 }
1432