Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * execParallel.c
4 : * Support routines for parallel execution.
5 : *
6 : * Portions Copyright (c) 1996-2017, 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/nodeBitmapHeapscan.h"
29 : #include "executor/nodeCustom.h"
30 : #include "executor/nodeForeignscan.h"
31 : #include "executor/nodeIndexscan.h"
32 : #include "executor/nodeIndexonlyscan.h"
33 : #include "executor/nodeSeqscan.h"
34 : #include "executor/nodeSort.h"
35 : #include "executor/tqueue.h"
36 : #include "nodes/nodeFuncs.h"
37 : #include "optimizer/planmain.h"
38 : #include "optimizer/planner.h"
39 : #include "storage/spin.h"
40 : #include "tcop/tcopprot.h"
41 : #include "utils/dsa.h"
42 : #include "utils/memutils.h"
43 : #include "utils/snapmgr.h"
44 : #include "pgstat.h"
45 :
46 : /*
47 : * Magic numbers for parallel executor communication. We use constants
48 : * greater than any 32-bit integer here so that values < 2^32 can be used
49 : * by individual parallel nodes to store their own state.
50 : */
51 : #define PARALLEL_KEY_EXECUTOR_FIXED UINT64CONST(0xE000000000000001)
52 : #define PARALLEL_KEY_PLANNEDSTMT UINT64CONST(0xE000000000000002)
53 : #define PARALLEL_KEY_PARAMS UINT64CONST(0xE000000000000003)
54 : #define PARALLEL_KEY_BUFFER_USAGE UINT64CONST(0xE000000000000004)
55 : #define PARALLEL_KEY_TUPLE_QUEUE UINT64CONST(0xE000000000000005)
56 : #define PARALLEL_KEY_INSTRUMENTATION UINT64CONST(0xE000000000000006)
57 : #define PARALLEL_KEY_DSA UINT64CONST(0xE000000000000007)
58 : #define PARALLEL_KEY_QUERY_TEXT UINT64CONST(0xE000000000000008)
59 :
60 : #define PARALLEL_TUPLE_QUEUE_SIZE 65536
61 :
62 : /*
63 : * Fixed-size random stuff that we need to pass to parallel workers.
64 : */
65 : typedef struct FixedParallelExecutorState
66 : {
67 : int64 tuples_needed; /* tuple bound, see ExecSetTupleBound */
68 : } FixedParallelExecutorState;
69 :
70 : /*
71 : * DSM structure for accumulating per-PlanState instrumentation.
72 : *
73 : * instrument_options: Same meaning here as in instrument.c.
74 : *
75 : * instrument_offset: Offset, relative to the start of this structure,
76 : * of the first Instrumentation object. This will depend on the length of
77 : * the plan_node_id array.
78 : *
79 : * num_workers: Number of workers.
80 : *
81 : * num_plan_nodes: Number of plan nodes.
82 : *
83 : * plan_node_id: Array of plan nodes for which we are gathering instrumentation
84 : * from parallel workers. The length of this array is given by num_plan_nodes.
85 : */
86 : struct SharedExecutorInstrumentation
87 : {
88 : int instrument_options;
89 : int instrument_offset;
90 : int num_workers;
91 : int num_plan_nodes;
92 : int plan_node_id[FLEXIBLE_ARRAY_MEMBER];
93 : /* array of num_plan_nodes * num_workers Instrumentation objects follows */
94 : };
95 : #define GetInstrumentationArray(sei) \
96 : (AssertVariableIsOfTypeMacro(sei, SharedExecutorInstrumentation *), \
97 : (Instrumentation *) (((char *) sei) + sei->instrument_offset))
98 :
99 : /* Context object for ExecParallelEstimate. */
100 : typedef struct ExecParallelEstimateContext
101 : {
102 : ParallelContext *pcxt;
103 : int nnodes;
104 : } ExecParallelEstimateContext;
105 :
106 : /* Context object for ExecParallelInitializeDSM. */
107 : typedef struct ExecParallelInitializeDSMContext
108 : {
109 : ParallelContext *pcxt;
110 : SharedExecutorInstrumentation *instrumentation;
111 : int nnodes;
112 : } ExecParallelInitializeDSMContext;
113 :
114 : /* Helper functions that run in the parallel leader. */
115 : static char *ExecSerializePlan(Plan *plan, EState *estate);
116 : static bool ExecParallelEstimate(PlanState *node,
117 : ExecParallelEstimateContext *e);
118 : static bool ExecParallelInitializeDSM(PlanState *node,
119 : ExecParallelInitializeDSMContext *d);
120 : static shm_mq_handle **ExecParallelSetupTupleQueues(ParallelContext *pcxt,
121 : bool reinitialize);
122 : static bool ExecParallelReInitializeDSM(PlanState *planstate,
123 : ParallelContext *pcxt);
124 : static bool ExecParallelRetrieveInstrumentation(PlanState *planstate,
125 : SharedExecutorInstrumentation *instrumentation);
126 :
127 : /* Helper function that runs in the parallel worker. */
128 : static DestReceiver *ExecParallelGetReceiver(dsm_segment *seg, shm_toc *toc);
129 :
130 : /*
131 : * Create a serialized representation of the plan to be sent to each worker.
132 : */
133 : static char *
134 17 : ExecSerializePlan(Plan *plan, EState *estate)
135 : {
136 : PlannedStmt *pstmt;
137 : ListCell *lc;
138 :
139 : /* We can't scribble on the original plan, so make a copy. */
140 17 : plan = copyObject(plan);
141 :
142 : /*
143 : * The worker will start its own copy of the executor, and that copy will
144 : * insert a junk filter if the toplevel node has any resjunk entries. We
145 : * don't want that to happen, because while resjunk columns shouldn't be
146 : * sent back to the user, here the tuples are coming back to another
147 : * backend which may very well need them. So mutate the target list
148 : * accordingly. This is sort of a hack; there might be better ways to do
149 : * this...
150 : */
151 51 : foreach(lc, plan->targetlist)
152 : {
153 34 : TargetEntry *tle = lfirst_node(TargetEntry, lc);
154 :
155 34 : tle->resjunk = false;
156 : }
157 :
158 : /*
159 : * Create a dummy PlannedStmt. Most of the fields don't need to be valid
160 : * for our purposes, but the worker will need at least a minimal
161 : * PlannedStmt to start the executor.
162 : */
163 17 : pstmt = makeNode(PlannedStmt);
164 17 : pstmt->commandType = CMD_SELECT;
165 17 : pstmt->queryId = 0;
166 17 : pstmt->hasReturning = false;
167 17 : pstmt->hasModifyingCTE = false;
168 17 : pstmt->canSetTag = true;
169 17 : pstmt->transientPlan = false;
170 17 : pstmt->dependsOnRole = false;
171 17 : pstmt->parallelModeNeeded = false;
172 17 : pstmt->planTree = plan;
173 17 : pstmt->rtable = estate->es_range_table;
174 17 : pstmt->resultRelations = NIL;
175 17 : pstmt->nonleafResultRelations = NIL;
176 :
177 : /*
178 : * Transfer only parallel-safe subplans, leaving a NULL "hole" in the list
179 : * for unsafe ones (so that the list indexes of the safe ones are
180 : * preserved). This positively ensures that the worker won't try to run,
181 : * or even do ExecInitNode on, an unsafe subplan. That's important to
182 : * protect, eg, non-parallel-aware FDWs from getting into trouble.
183 : */
184 17 : pstmt->subplans = NIL;
185 18 : foreach(lc, estate->es_plannedstmt->subplans)
186 : {
187 1 : Plan *subplan = (Plan *) lfirst(lc);
188 :
189 1 : if (subplan && !subplan->parallel_safe)
190 0 : subplan = NULL;
191 1 : pstmt->subplans = lappend(pstmt->subplans, subplan);
192 : }
193 :
194 17 : pstmt->rewindPlanIDs = NULL;
195 17 : pstmt->rowMarks = NIL;
196 17 : pstmt->relationOids = NIL;
197 17 : pstmt->invalItems = NIL; /* workers can't replan anyway... */
198 17 : pstmt->nParamExec = estate->es_plannedstmt->nParamExec;
199 17 : pstmt->utilityStmt = NULL;
200 17 : pstmt->stmt_location = -1;
201 17 : pstmt->stmt_len = -1;
202 :
203 : /* Return serialized copy of our dummy PlannedStmt. */
204 17 : return nodeToString(pstmt);
205 : }
206 :
207 : /*
208 : * Parallel-aware plan nodes (and occasionally others) may need some state
209 : * which is shared across all parallel workers. Before we size the DSM, give
210 : * them a chance to call shm_toc_estimate_chunk or shm_toc_estimate_keys on
211 : * &pcxt->estimator.
212 : *
213 : * While we're at it, count the number of PlanState nodes in the tree, so
214 : * we know how many SharedPlanStateInstrumentation structures we need.
215 : */
216 : static bool
217 46 : ExecParallelEstimate(PlanState *planstate, ExecParallelEstimateContext *e)
218 : {
219 46 : if (planstate == NULL)
220 0 : return false;
221 :
222 : /* Count this node. */
223 46 : e->nnodes++;
224 :
225 46 : switch (nodeTag(planstate))
226 : {
227 : case T_SeqScanState:
228 14 : if (planstate->plan->parallel_aware)
229 13 : ExecSeqScanEstimate((SeqScanState *) planstate,
230 : e->pcxt);
231 14 : break;
232 : case T_IndexScanState:
233 3 : if (planstate->plan->parallel_aware)
234 2 : ExecIndexScanEstimate((IndexScanState *) planstate,
235 : e->pcxt);
236 3 : break;
237 : case T_IndexOnlyScanState:
238 4 : if (planstate->plan->parallel_aware)
239 3 : ExecIndexOnlyScanEstimate((IndexOnlyScanState *) planstate,
240 : e->pcxt);
241 4 : break;
242 : case T_ForeignScanState:
243 0 : if (planstate->plan->parallel_aware)
244 0 : ExecForeignScanEstimate((ForeignScanState *) planstate,
245 : e->pcxt);
246 0 : break;
247 : case T_CustomScanState:
248 0 : if (planstate->plan->parallel_aware)
249 0 : ExecCustomScanEstimate((CustomScanState *) planstate,
250 : e->pcxt);
251 0 : break;
252 : case T_BitmapHeapScanState:
253 2 : if (planstate->plan->parallel_aware)
254 2 : ExecBitmapHeapEstimate((BitmapHeapScanState *) planstate,
255 : e->pcxt);
256 2 : break;
257 : case T_SortState:
258 : /* even when not parallel-aware */
259 5 : ExecSortEstimate((SortState *) planstate, e->pcxt);
260 5 : break;
261 :
262 : default:
263 18 : break;
264 : }
265 :
266 46 : return planstate_tree_walker(planstate, ExecParallelEstimate, e);
267 : }
268 :
269 : /*
270 : * Initialize the dynamic shared memory segment that will be used to control
271 : * parallel execution.
272 : */
273 : static bool
274 46 : ExecParallelInitializeDSM(PlanState *planstate,
275 : ExecParallelInitializeDSMContext *d)
276 : {
277 46 : if (planstate == NULL)
278 0 : return false;
279 :
280 : /* If instrumentation is enabled, initialize slot for this node. */
281 46 : if (d->instrumentation != NULL)
282 2 : d->instrumentation->plan_node_id[d->nnodes] =
283 1 : planstate->plan->plan_node_id;
284 :
285 : /* Count this node. */
286 46 : d->nnodes++;
287 :
288 : /*
289 : * Call initializers for DSM-using plan nodes.
290 : *
291 : * Most plan nodes won't do anything here, but plan nodes that allocated
292 : * DSM may need to initialize shared state in the DSM before parallel
293 : * workers are launched. They can allocate the space they previously
294 : * estimated using shm_toc_allocate, and add the keys they previously
295 : * estimated using shm_toc_insert, in each case targeting pcxt->toc.
296 : */
297 46 : switch (nodeTag(planstate))
298 : {
299 : case T_SeqScanState:
300 14 : if (planstate->plan->parallel_aware)
301 13 : ExecSeqScanInitializeDSM((SeqScanState *) planstate,
302 : d->pcxt);
303 14 : break;
304 : case T_IndexScanState:
305 3 : if (planstate->plan->parallel_aware)
306 2 : ExecIndexScanInitializeDSM((IndexScanState *) planstate,
307 : d->pcxt);
308 3 : break;
309 : case T_IndexOnlyScanState:
310 4 : if (planstate->plan->parallel_aware)
311 3 : ExecIndexOnlyScanInitializeDSM((IndexOnlyScanState *) planstate,
312 : d->pcxt);
313 4 : break;
314 : case T_ForeignScanState:
315 0 : if (planstate->plan->parallel_aware)
316 0 : ExecForeignScanInitializeDSM((ForeignScanState *) planstate,
317 : d->pcxt);
318 0 : break;
319 : case T_CustomScanState:
320 0 : if (planstate->plan->parallel_aware)
321 0 : ExecCustomScanInitializeDSM((CustomScanState *) planstate,
322 : d->pcxt);
323 0 : break;
324 : case T_BitmapHeapScanState:
325 2 : if (planstate->plan->parallel_aware)
326 2 : ExecBitmapHeapInitializeDSM((BitmapHeapScanState *) planstate,
327 : d->pcxt);
328 2 : break;
329 : case T_SortState:
330 : /* even when not parallel-aware */
331 5 : ExecSortInitializeDSM((SortState *) planstate, d->pcxt);
332 5 : break;
333 :
334 : default:
335 18 : break;
336 : }
337 :
338 46 : return planstate_tree_walker(planstate, ExecParallelInitializeDSM, d);
339 : }
340 :
341 : /*
342 : * It sets up the response queues for backend workers to return tuples
343 : * to the main backend and start the workers.
344 : */
345 : static shm_mq_handle **
346 32 : ExecParallelSetupTupleQueues(ParallelContext *pcxt, bool reinitialize)
347 : {
348 : shm_mq_handle **responseq;
349 : char *tqueuespace;
350 : int i;
351 :
352 : /* Skip this if no workers. */
353 32 : if (pcxt->nworkers == 0)
354 0 : return NULL;
355 :
356 : /* Allocate memory for shared memory queue handles. */
357 32 : responseq = (shm_mq_handle **)
358 32 : palloc(pcxt->nworkers * sizeof(shm_mq_handle *));
359 :
360 : /*
361 : * If not reinitializing, allocate space from the DSM for the queues;
362 : * otherwise, find the already allocated space.
363 : */
364 32 : if (!reinitialize)
365 17 : tqueuespace =
366 17 : shm_toc_allocate(pcxt->toc,
367 : mul_size(PARALLEL_TUPLE_QUEUE_SIZE,
368 17 : pcxt->nworkers));
369 : else
370 15 : tqueuespace = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_TUPLE_QUEUE, false);
371 :
372 : /* Create the queues, and become the receiver for each. */
373 151 : for (i = 0; i < pcxt->nworkers; ++i)
374 : {
375 : shm_mq *mq;
376 :
377 119 : mq = shm_mq_create(tqueuespace +
378 119 : ((Size) i) * PARALLEL_TUPLE_QUEUE_SIZE,
379 : (Size) PARALLEL_TUPLE_QUEUE_SIZE);
380 :
381 119 : shm_mq_set_receiver(mq, MyProc);
382 119 : responseq[i] = shm_mq_attach(mq, pcxt->seg, NULL);
383 : }
384 :
385 : /* Add array of queues to shm_toc, so others can find it. */
386 32 : if (!reinitialize)
387 17 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_TUPLE_QUEUE, tqueuespace);
388 :
389 : /* Return array of handles. */
390 32 : return responseq;
391 : }
392 :
393 : /*
394 : * Sets up the required infrastructure for backend workers to perform
395 : * execution and return results to the main backend.
396 : */
397 : ParallelExecutorInfo *
398 17 : ExecInitParallelPlan(PlanState *planstate, EState *estate, int nworkers,
399 : int64 tuples_needed)
400 : {
401 : ParallelExecutorInfo *pei;
402 : ParallelContext *pcxt;
403 : ExecParallelEstimateContext e;
404 : ExecParallelInitializeDSMContext d;
405 : FixedParallelExecutorState *fpes;
406 : char *pstmt_data;
407 : char *pstmt_space;
408 : char *param_space;
409 : BufferUsage *bufusage_space;
410 17 : SharedExecutorInstrumentation *instrumentation = NULL;
411 : int pstmt_len;
412 : int param_len;
413 17 : int instrumentation_len = 0;
414 17 : int instrument_offset = 0;
415 17 : Size dsa_minsize = dsa_minimum_size();
416 : char *query_string;
417 : int query_len;
418 :
419 : /* Allocate object for return value. */
420 17 : pei = palloc0(sizeof(ParallelExecutorInfo));
421 17 : pei->finished = false;
422 17 : pei->planstate = planstate;
423 :
424 : /* Fix up and serialize plan to be sent to workers. */
425 17 : pstmt_data = ExecSerializePlan(planstate->plan, estate);
426 :
427 : /* Create a parallel context. */
428 17 : pcxt = CreateParallelContext("postgres", "ParallelQueryMain", nworkers);
429 17 : pei->pcxt = pcxt;
430 :
431 : /*
432 : * Before telling the parallel context to create a dynamic shared memory
433 : * segment, we need to figure out how big it should be. Estimate space
434 : * for the various things we need to store.
435 : */
436 :
437 : /* Estimate space for fixed-size state. */
438 17 : shm_toc_estimate_chunk(&pcxt->estimator,
439 : sizeof(FixedParallelExecutorState));
440 17 : shm_toc_estimate_keys(&pcxt->estimator, 1);
441 :
442 : /* Estimate space for query text. */
443 17 : query_len = strlen(estate->es_sourceText);
444 17 : shm_toc_estimate_chunk(&pcxt->estimator, query_len);
445 17 : shm_toc_estimate_keys(&pcxt->estimator, 1);
446 :
447 : /* Estimate space for serialized PlannedStmt. */
448 17 : pstmt_len = strlen(pstmt_data) + 1;
449 17 : shm_toc_estimate_chunk(&pcxt->estimator, pstmt_len);
450 17 : shm_toc_estimate_keys(&pcxt->estimator, 1);
451 :
452 : /* Estimate space for serialized ParamListInfo. */
453 17 : param_len = EstimateParamListSpace(estate->es_param_list_info);
454 17 : shm_toc_estimate_chunk(&pcxt->estimator, param_len);
455 17 : shm_toc_estimate_keys(&pcxt->estimator, 1);
456 :
457 : /*
458 : * Estimate space for BufferUsage.
459 : *
460 : * If EXPLAIN is not in use and there are no extensions loaded that care,
461 : * we could skip this. But we have no way of knowing whether anyone's
462 : * looking at pgBufferUsage, so do it unconditionally.
463 : */
464 17 : shm_toc_estimate_chunk(&pcxt->estimator,
465 : mul_size(sizeof(BufferUsage), pcxt->nworkers));
466 17 : shm_toc_estimate_keys(&pcxt->estimator, 1);
467 :
468 : /* Estimate space for tuple queues. */
469 17 : shm_toc_estimate_chunk(&pcxt->estimator,
470 : mul_size(PARALLEL_TUPLE_QUEUE_SIZE, pcxt->nworkers));
471 17 : shm_toc_estimate_keys(&pcxt->estimator, 1);
472 :
473 : /*
474 : * Give parallel-aware nodes a chance to add to the estimates, and get a
475 : * count of how many PlanState nodes there are.
476 : */
477 17 : e.pcxt = pcxt;
478 17 : e.nnodes = 0;
479 17 : ExecParallelEstimate(planstate, &e);
480 :
481 : /* Estimate space for instrumentation, if required. */
482 17 : if (estate->es_instrument)
483 : {
484 1 : instrumentation_len =
485 1 : offsetof(SharedExecutorInstrumentation, plan_node_id) +
486 1 : sizeof(int) * e.nnodes;
487 1 : instrumentation_len = MAXALIGN(instrumentation_len);
488 1 : instrument_offset = instrumentation_len;
489 1 : instrumentation_len +=
490 2 : mul_size(sizeof(Instrumentation),
491 1 : mul_size(e.nnodes, nworkers));
492 1 : shm_toc_estimate_chunk(&pcxt->estimator, instrumentation_len);
493 1 : shm_toc_estimate_keys(&pcxt->estimator, 1);
494 : }
495 :
496 : /* Estimate space for DSA area. */
497 17 : shm_toc_estimate_chunk(&pcxt->estimator, dsa_minsize);
498 17 : shm_toc_estimate_keys(&pcxt->estimator, 1);
499 :
500 : /* Everyone's had a chance to ask for space, so now create the DSM. */
501 17 : InitializeParallelDSM(pcxt);
502 :
503 : /*
504 : * OK, now we have a dynamic shared memory segment, and it should be big
505 : * enough to store all of the data we estimated we would want to put into
506 : * it, plus whatever general stuff (not specifically executor-related) the
507 : * ParallelContext itself needs to store there. None of the space we
508 : * asked for has been allocated or initialized yet, though, so do that.
509 : */
510 :
511 : /* Store fixed-size state. */
512 17 : fpes = shm_toc_allocate(pcxt->toc, sizeof(FixedParallelExecutorState));
513 17 : fpes->tuples_needed = tuples_needed;
514 17 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_EXECUTOR_FIXED, fpes);
515 :
516 : /* Store query string */
517 17 : query_string = shm_toc_allocate(pcxt->toc, query_len);
518 17 : memcpy(query_string, estate->es_sourceText, query_len);
519 17 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_QUERY_TEXT, query_string);
520 :
521 : /* Store serialized PlannedStmt. */
522 17 : pstmt_space = shm_toc_allocate(pcxt->toc, pstmt_len);
523 17 : memcpy(pstmt_space, pstmt_data, pstmt_len);
524 17 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_PLANNEDSTMT, pstmt_space);
525 :
526 : /* Store serialized ParamListInfo. */
527 17 : param_space = shm_toc_allocate(pcxt->toc, param_len);
528 17 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_PARAMS, param_space);
529 17 : SerializeParamList(estate->es_param_list_info, ¶m_space);
530 :
531 : /* Allocate space for each worker's BufferUsage; no need to initialize. */
532 17 : bufusage_space = shm_toc_allocate(pcxt->toc,
533 17 : mul_size(sizeof(BufferUsage), pcxt->nworkers));
534 17 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_BUFFER_USAGE, bufusage_space);
535 17 : pei->buffer_usage = bufusage_space;
536 :
537 : /* Set up the tuple queues that the workers will write into. */
538 17 : pei->tqueue = ExecParallelSetupTupleQueues(pcxt, false);
539 :
540 : /* We don't need the TupleQueueReaders yet, though. */
541 17 : pei->reader = NULL;
542 :
543 : /*
544 : * If instrumentation options were supplied, allocate space for the data.
545 : * It only gets partially initialized here; the rest happens during
546 : * ExecParallelInitializeDSM.
547 : */
548 17 : if (estate->es_instrument)
549 : {
550 : Instrumentation *instrument;
551 : int i;
552 :
553 1 : instrumentation = shm_toc_allocate(pcxt->toc, instrumentation_len);
554 1 : instrumentation->instrument_options = estate->es_instrument;
555 1 : instrumentation->instrument_offset = instrument_offset;
556 1 : instrumentation->num_workers = nworkers;
557 1 : instrumentation->num_plan_nodes = e.nnodes;
558 1 : instrument = GetInstrumentationArray(instrumentation);
559 5 : for (i = 0; i < nworkers * e.nnodes; ++i)
560 4 : InstrInit(&instrument[i], estate->es_instrument);
561 1 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_INSTRUMENTATION,
562 : instrumentation);
563 1 : pei->instrumentation = instrumentation;
564 : }
565 :
566 : /*
567 : * Create a DSA area that can be used by the leader and all workers.
568 : * (However, if we failed to create a DSM and are using private memory
569 : * instead, then skip this.)
570 : */
571 17 : if (pcxt->seg != NULL)
572 : {
573 : char *area_space;
574 :
575 17 : area_space = shm_toc_allocate(pcxt->toc, dsa_minsize);
576 17 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_DSA, area_space);
577 17 : pei->area = dsa_create_in_place(area_space, dsa_minsize,
578 : LWTRANCHE_PARALLEL_QUERY_DSA,
579 : pcxt->seg);
580 : }
581 :
582 : /*
583 : * Make the area available to executor nodes running in the leader. See
584 : * also ParallelQueryMain which makes it available to workers.
585 : */
586 17 : estate->es_query_dsa = pei->area;
587 :
588 : /*
589 : * Give parallel-aware nodes a chance to initialize their shared data.
590 : * This also initializes the elements of instrumentation->ps_instrument,
591 : * if it exists.
592 : */
593 17 : d.pcxt = pcxt;
594 17 : d.instrumentation = instrumentation;
595 17 : d.nnodes = 0;
596 17 : ExecParallelInitializeDSM(planstate, &d);
597 :
598 : /*
599 : * Make sure that the world hasn't shifted under our feet. This could
600 : * probably just be an Assert(), but let's be conservative for now.
601 : */
602 17 : if (e.nnodes != d.nnodes)
603 0 : elog(ERROR, "inconsistent count of PlanState nodes");
604 :
605 : /* OK, we're ready to rock and roll. */
606 17 : return pei;
607 : }
608 :
609 : /*
610 : * Set up tuple queue readers to read the results of a parallel subplan.
611 : * All the workers are expected to return tuples matching tupDesc.
612 : *
613 : * This is separate from ExecInitParallelPlan() because we can launch the
614 : * worker processes and let them start doing something before we do this.
615 : */
616 : void
617 31 : ExecParallelCreateReaders(ParallelExecutorInfo *pei,
618 : TupleDesc tupDesc)
619 : {
620 31 : int nworkers = pei->pcxt->nworkers_launched;
621 : int i;
622 :
623 31 : Assert(pei->reader == NULL);
624 :
625 31 : if (nworkers > 0)
626 : {
627 31 : pei->reader = (TupleQueueReader **)
628 31 : palloc(nworkers * sizeof(TupleQueueReader *));
629 :
630 146 : for (i = 0; i < nworkers; i++)
631 : {
632 115 : shm_mq_set_handle(pei->tqueue[i],
633 115 : pei->pcxt->worker[i].bgwhandle);
634 115 : pei->reader[i] = CreateTupleQueueReader(pei->tqueue[i],
635 : tupDesc);
636 : }
637 : }
638 31 : }
639 :
640 : /*
641 : * Re-initialize the parallel executor shared memory state before launching
642 : * a fresh batch of workers.
643 : */
644 : void
645 15 : ExecParallelReinitialize(PlanState *planstate,
646 : ParallelExecutorInfo *pei)
647 : {
648 : /* Old workers must already be shut down */
649 15 : Assert(pei->finished);
650 :
651 15 : ReinitializeParallelDSM(pei->pcxt);
652 15 : pei->tqueue = ExecParallelSetupTupleQueues(pei->pcxt, true);
653 15 : pei->reader = NULL;
654 15 : pei->finished = false;
655 :
656 : /* Traverse plan tree and let each child node reset associated state. */
657 15 : ExecParallelReInitializeDSM(planstate, pei->pcxt);
658 15 : }
659 :
660 : /*
661 : * Traverse plan tree to reinitialize per-node dynamic shared memory state
662 : */
663 : static bool
664 32 : ExecParallelReInitializeDSM(PlanState *planstate,
665 : ParallelContext *pcxt)
666 : {
667 32 : if (planstate == NULL)
668 0 : return false;
669 :
670 : /*
671 : * Call reinitializers for DSM-using plan nodes.
672 : */
673 32 : switch (nodeTag(planstate))
674 : {
675 : case T_SeqScanState:
676 2 : if (planstate->plan->parallel_aware)
677 2 : ExecSeqScanReInitializeDSM((SeqScanState *) planstate,
678 : pcxt);
679 2 : break;
680 : case T_IndexScanState:
681 2 : if (planstate->plan->parallel_aware)
682 2 : ExecIndexScanReInitializeDSM((IndexScanState *) planstate,
683 : pcxt);
684 2 : break;
685 : case T_IndexOnlyScanState:
686 2 : if (planstate->plan->parallel_aware)
687 2 : ExecIndexOnlyScanReInitializeDSM((IndexOnlyScanState *) planstate,
688 : pcxt);
689 2 : break;
690 : case T_ForeignScanState:
691 0 : if (planstate->plan->parallel_aware)
692 0 : ExecForeignScanReInitializeDSM((ForeignScanState *) planstate,
693 : pcxt);
694 0 : break;
695 : case T_CustomScanState:
696 0 : if (planstate->plan->parallel_aware)
697 0 : ExecCustomScanReInitializeDSM((CustomScanState *) planstate,
698 : pcxt);
699 0 : break;
700 : case T_BitmapHeapScanState:
701 9 : if (planstate->plan->parallel_aware)
702 9 : ExecBitmapHeapReInitializeDSM((BitmapHeapScanState *) planstate,
703 : pcxt);
704 9 : break;
705 : case T_SortState:
706 : /* even when not parallel-aware */
707 2 : ExecSortReInitializeDSM((SortState *) planstate, pcxt);
708 2 : break;
709 :
710 : default:
711 15 : break;
712 : }
713 :
714 32 : return planstate_tree_walker(planstate, ExecParallelReInitializeDSM, pcxt);
715 : }
716 :
717 : /*
718 : * Copy instrumentation information about this node and its descendants from
719 : * dynamic shared memory.
720 : */
721 : static bool
722 1 : ExecParallelRetrieveInstrumentation(PlanState *planstate,
723 : SharedExecutorInstrumentation *instrumentation)
724 : {
725 : Instrumentation *instrument;
726 : int i;
727 : int n;
728 : int ibytes;
729 1 : int plan_node_id = planstate->plan->plan_node_id;
730 : MemoryContext oldcontext;
731 :
732 : /* Find the instrumentation for this node. */
733 1 : for (i = 0; i < instrumentation->num_plan_nodes; ++i)
734 1 : if (instrumentation->plan_node_id[i] == plan_node_id)
735 1 : break;
736 1 : if (i >= instrumentation->num_plan_nodes)
737 0 : elog(ERROR, "plan node %d not found", plan_node_id);
738 :
739 : /* Accumulate the statistics from all workers. */
740 1 : instrument = GetInstrumentationArray(instrumentation);
741 1 : instrument += i * instrumentation->num_workers;
742 5 : for (n = 0; n < instrumentation->num_workers; ++n)
743 4 : InstrAggNode(planstate->instrument, &instrument[n]);
744 :
745 : /*
746 : * Also store the per-worker detail.
747 : *
748 : * Worker instrumentation should be allocated in the same context as the
749 : * regular instrumentation information, which is the per-query context.
750 : * Switch into per-query memory context.
751 : */
752 1 : oldcontext = MemoryContextSwitchTo(planstate->state->es_query_cxt);
753 1 : ibytes = mul_size(instrumentation->num_workers, sizeof(Instrumentation));
754 1 : planstate->worker_instrument =
755 1 : palloc(ibytes + offsetof(WorkerInstrumentation, instrument));
756 1 : MemoryContextSwitchTo(oldcontext);
757 :
758 1 : planstate->worker_instrument->num_workers = instrumentation->num_workers;
759 1 : memcpy(&planstate->worker_instrument->instrument, instrument, ibytes);
760 :
761 : /*
762 : * Perform any node-type-specific work that needs to be done. Currently,
763 : * only Sort nodes need to do anything here.
764 : */
765 1 : if (IsA(planstate, SortState))
766 0 : ExecSortRetrieveInstrumentation((SortState *) planstate);
767 :
768 1 : return planstate_tree_walker(planstate, ExecParallelRetrieveInstrumentation,
769 : instrumentation);
770 : }
771 :
772 : /*
773 : * Finish parallel execution. We wait for parallel workers to finish, and
774 : * accumulate their buffer usage and instrumentation.
775 : */
776 : void
777 31 : ExecParallelFinish(ParallelExecutorInfo *pei)
778 : {
779 31 : int nworkers = pei->pcxt->nworkers_launched;
780 : int i;
781 :
782 : /* Make this be a no-op if called twice in a row. */
783 31 : if (pei->finished)
784 31 : return;
785 :
786 : /*
787 : * Detach from tuple queues ASAP, so that any still-active workers will
788 : * notice that no further results are wanted.
789 : */
790 31 : if (pei->tqueue != NULL)
791 : {
792 145 : for (i = 0; i < nworkers; i++)
793 114 : shm_mq_detach(pei->tqueue[i]);
794 31 : pfree(pei->tqueue);
795 31 : pei->tqueue = NULL;
796 : }
797 :
798 : /*
799 : * While we're waiting for the workers to finish, let's get rid of the
800 : * tuple queue readers. (Any other local cleanup could be done here too.)
801 : */
802 31 : if (pei->reader != NULL)
803 : {
804 144 : for (i = 0; i < nworkers; i++)
805 114 : DestroyTupleQueueReader(pei->reader[i]);
806 30 : pfree(pei->reader);
807 30 : pei->reader = NULL;
808 : }
809 :
810 : /* Now wait for the workers to finish. */
811 31 : WaitForParallelWorkersToFinish(pei->pcxt);
812 :
813 : /*
814 : * Next, accumulate buffer usage. (This must wait for the workers to
815 : * finish, or we might get incomplete data.)
816 : */
817 145 : for (i = 0; i < nworkers; i++)
818 114 : InstrAccumParallelQuery(&pei->buffer_usage[i]);
819 :
820 : /* Finally, accumulate instrumentation, if any. */
821 31 : if (pei->instrumentation)
822 1 : ExecParallelRetrieveInstrumentation(pei->planstate,
823 : pei->instrumentation);
824 :
825 31 : pei->finished = true;
826 : }
827 :
828 : /*
829 : * Clean up whatever ParallelExecutorInfo resources still exist after
830 : * ExecParallelFinish. We separate these routines because someone might
831 : * want to examine the contents of the DSM after ExecParallelFinish and
832 : * before calling this routine.
833 : */
834 : void
835 16 : ExecParallelCleanup(ParallelExecutorInfo *pei)
836 : {
837 16 : if (pei->area != NULL)
838 : {
839 16 : dsa_detach(pei->area);
840 16 : pei->area = NULL;
841 : }
842 16 : if (pei->pcxt != NULL)
843 : {
844 16 : DestroyParallelContext(pei->pcxt);
845 16 : pei->pcxt = NULL;
846 : }
847 16 : pfree(pei);
848 16 : }
849 :
850 : /*
851 : * Create a DestReceiver to write tuples we produce to the shm_mq designated
852 : * for that purpose.
853 : */
854 : static DestReceiver *
855 115 : ExecParallelGetReceiver(dsm_segment *seg, shm_toc *toc)
856 : {
857 : char *mqspace;
858 : shm_mq *mq;
859 :
860 115 : mqspace = shm_toc_lookup(toc, PARALLEL_KEY_TUPLE_QUEUE, false);
861 115 : mqspace += ParallelWorkerNumber * PARALLEL_TUPLE_QUEUE_SIZE;
862 115 : mq = (shm_mq *) mqspace;
863 115 : shm_mq_set_sender(mq, MyProc);
864 115 : return CreateTupleQueueDestReceiver(shm_mq_attach(mq, seg, NULL));
865 : }
866 :
867 : /*
868 : * Create a QueryDesc for the PlannedStmt we are to execute, and return it.
869 : */
870 : static QueryDesc *
871 115 : ExecParallelGetQueryDesc(shm_toc *toc, DestReceiver *receiver,
872 : int instrument_options)
873 : {
874 : char *pstmtspace;
875 : char *paramspace;
876 : PlannedStmt *pstmt;
877 : ParamListInfo paramLI;
878 : char *queryString;
879 :
880 : /* Get the query string from shared memory */
881 115 : queryString = shm_toc_lookup(toc, PARALLEL_KEY_QUERY_TEXT, false);
882 :
883 : /* Reconstruct leader-supplied PlannedStmt. */
884 115 : pstmtspace = shm_toc_lookup(toc, PARALLEL_KEY_PLANNEDSTMT, false);
885 115 : pstmt = (PlannedStmt *) stringToNode(pstmtspace);
886 :
887 : /* Reconstruct ParamListInfo. */
888 115 : paramspace = shm_toc_lookup(toc, PARALLEL_KEY_PARAMS, false);
889 115 : paramLI = RestoreParamList(¶mspace);
890 :
891 : /*
892 : * Create a QueryDesc for the query.
893 : *
894 : * It's not obvious how to obtain the query string from here; and even if
895 : * we could copying it would take more cycles than not copying it. But
896 : * it's a bit unsatisfying to just use a dummy string here, so consider
897 : * revising this someday.
898 : */
899 115 : return CreateQueryDesc(pstmt,
900 : queryString,
901 : GetActiveSnapshot(), InvalidSnapshot,
902 : receiver, paramLI, NULL, instrument_options);
903 : }
904 :
905 : /*
906 : * Copy instrumentation information from this node and its descendants into
907 : * dynamic shared memory, so that the parallel leader can retrieve it.
908 : */
909 : static bool
910 4 : ExecParallelReportInstrumentation(PlanState *planstate,
911 : SharedExecutorInstrumentation *instrumentation)
912 : {
913 : int i;
914 4 : int plan_node_id = planstate->plan->plan_node_id;
915 : Instrumentation *instrument;
916 :
917 4 : InstrEndLoop(planstate->instrument);
918 :
919 : /*
920 : * If we shuffled the plan_node_id values in ps_instrument into sorted
921 : * order, we could use binary search here. This might matter someday if
922 : * we're pushing down sufficiently large plan trees. For now, do it the
923 : * slow, dumb way.
924 : */
925 4 : for (i = 0; i < instrumentation->num_plan_nodes; ++i)
926 4 : if (instrumentation->plan_node_id[i] == plan_node_id)
927 4 : break;
928 4 : if (i >= instrumentation->num_plan_nodes)
929 0 : elog(ERROR, "plan node %d not found", plan_node_id);
930 :
931 : /*
932 : * Add our statistics to the per-node, per-worker totals. It's possible
933 : * that this could happen more than once if we relaunched workers.
934 : */
935 4 : instrument = GetInstrumentationArray(instrumentation);
936 4 : instrument += i * instrumentation->num_workers;
937 4 : Assert(IsParallelWorker());
938 4 : Assert(ParallelWorkerNumber < instrumentation->num_workers);
939 4 : InstrAggNode(&instrument[ParallelWorkerNumber], planstate->instrument);
940 :
941 4 : return planstate_tree_walker(planstate, ExecParallelReportInstrumentation,
942 : instrumentation);
943 : }
944 :
945 : /*
946 : * Initialize the PlanState and its descendants with the information
947 : * retrieved from shared memory. This has to be done once the PlanState
948 : * is allocated and initialized by executor; that is, after ExecutorStart().
949 : */
950 : static bool
951 265 : ExecParallelInitializeWorker(PlanState *planstate, shm_toc *toc)
952 : {
953 265 : if (planstate == NULL)
954 0 : return false;
955 :
956 265 : switch (nodeTag(planstate))
957 : {
958 : case T_SeqScanState:
959 42 : if (planstate->plan->parallel_aware)
960 38 : ExecSeqScanInitializeWorker((SeqScanState *) planstate, toc);
961 42 : break;
962 : case T_IndexScanState:
963 17 : if (planstate->plan->parallel_aware)
964 16 : ExecIndexScanInitializeWorker((IndexScanState *) planstate, toc);
965 17 : break;
966 : case T_IndexOnlyScanState:
967 24 : if (planstate->plan->parallel_aware)
968 20 : ExecIndexOnlyScanInitializeWorker((IndexOnlyScanState *) planstate, toc);
969 24 : break;
970 : case T_ForeignScanState:
971 0 : if (planstate->plan->parallel_aware)
972 0 : ExecForeignScanInitializeWorker((ForeignScanState *) planstate,
973 : toc);
974 0 : break;
975 : case T_CustomScanState:
976 0 : if (planstate->plan->parallel_aware)
977 0 : ExecCustomScanInitializeWorker((CustomScanState *) planstate,
978 : toc);
979 0 : break;
980 : case T_BitmapHeapScanState:
981 44 : if (planstate->plan->parallel_aware)
982 44 : ExecBitmapHeapInitializeWorker((BitmapHeapScanState *) planstate, toc);
983 44 : break;
984 : case T_SortState:
985 : /* even when not parallel-aware */
986 21 : ExecSortInitializeWorker((SortState *) planstate, toc);
987 21 : break;
988 :
989 : default:
990 117 : break;
991 : }
992 :
993 265 : return planstate_tree_walker(planstate, ExecParallelInitializeWorker, toc);
994 : }
995 :
996 : /*
997 : * Main entrypoint for parallel query worker processes.
998 : *
999 : * We reach this function from ParallelWorkerMain, so the setup necessary to
1000 : * create a sensible parallel environment has already been done;
1001 : * ParallelWorkerMain worries about stuff like the transaction state, combo
1002 : * CID mappings, and GUC values, so we don't need to deal with any of that
1003 : * here.
1004 : *
1005 : * Our job is to deal with concerns specific to the executor. The parallel
1006 : * group leader will have stored a serialized PlannedStmt, and it's our job
1007 : * to execute that plan and write the resulting tuples to the appropriate
1008 : * tuple queue. Various bits of supporting information that we need in order
1009 : * to do this are also stored in the dsm_segment and can be accessed through
1010 : * the shm_toc.
1011 : */
1012 : void
1013 115 : ParallelQueryMain(dsm_segment *seg, shm_toc *toc)
1014 : {
1015 : FixedParallelExecutorState *fpes;
1016 : BufferUsage *buffer_usage;
1017 : DestReceiver *receiver;
1018 : QueryDesc *queryDesc;
1019 : SharedExecutorInstrumentation *instrumentation;
1020 115 : int instrument_options = 0;
1021 : void *area_space;
1022 : dsa_area *area;
1023 :
1024 : /* Get fixed-size state. */
1025 115 : fpes = shm_toc_lookup(toc, PARALLEL_KEY_EXECUTOR_FIXED, false);
1026 :
1027 : /* Set up DestReceiver, SharedExecutorInstrumentation, and QueryDesc. */
1028 115 : receiver = ExecParallelGetReceiver(seg, toc);
1029 115 : instrumentation = shm_toc_lookup(toc, PARALLEL_KEY_INSTRUMENTATION, true);
1030 115 : if (instrumentation != NULL)
1031 4 : instrument_options = instrumentation->instrument_options;
1032 115 : queryDesc = ExecParallelGetQueryDesc(toc, receiver, instrument_options);
1033 :
1034 : /* Setting debug_query_string for individual workers */
1035 115 : debug_query_string = queryDesc->sourceText;
1036 :
1037 : /* Report workers' query for monitoring purposes */
1038 115 : pgstat_report_activity(STATE_RUNNING, debug_query_string);
1039 :
1040 : /* Prepare to track buffer usage during query execution. */
1041 115 : InstrStartParallelQuery();
1042 :
1043 : /* Attach to the dynamic shared memory area. */
1044 115 : area_space = shm_toc_lookup(toc, PARALLEL_KEY_DSA, false);
1045 115 : area = dsa_attach_in_place(area_space, seg);
1046 :
1047 : /* Start up the executor */
1048 115 : ExecutorStart(queryDesc, 0);
1049 :
1050 : /* Special executor initialization steps for parallel workers */
1051 115 : queryDesc->planstate->state->es_query_dsa = area;
1052 115 : ExecParallelInitializeWorker(queryDesc->planstate, toc);
1053 :
1054 : /* Pass down any tuple bound */
1055 115 : ExecSetTupleBound(fpes->tuples_needed, queryDesc->planstate);
1056 :
1057 : /*
1058 : * Run the plan. If we specified a tuple bound, be careful not to demand
1059 : * more tuples than that.
1060 : */
1061 115 : ExecutorRun(queryDesc,
1062 : ForwardScanDirection,
1063 115 : fpes->tuples_needed < 0 ? (int64) 0 : fpes->tuples_needed,
1064 : true);
1065 :
1066 : /* Shut down the executor */
1067 114 : ExecutorFinish(queryDesc);
1068 :
1069 : /* Report buffer usage during parallel execution. */
1070 114 : buffer_usage = shm_toc_lookup(toc, PARALLEL_KEY_BUFFER_USAGE, false);
1071 114 : InstrEndParallelQuery(&buffer_usage[ParallelWorkerNumber]);
1072 :
1073 : /* Report instrumentation data if any instrumentation options are set. */
1074 114 : if (instrumentation != NULL)
1075 4 : ExecParallelReportInstrumentation(queryDesc->planstate,
1076 : instrumentation);
1077 :
1078 : /* Must do this after capturing instrumentation. */
1079 114 : ExecutorEnd(queryDesc);
1080 :
1081 : /* Cleanup. */
1082 114 : dsa_detach(area);
1083 114 : FreeQueryDesc(queryDesc);
1084 114 : (*receiver->rDestroy) (receiver);
1085 114 : }
|
