Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * nodeGatherMerge.c
4 : * Scan a plan in multiple workers, and do order-preserving merge.
5 : *
6 : * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : * IDENTIFICATION
10 : * src/backend/executor/nodeGatherMerge.c
11 : *
12 : *-------------------------------------------------------------------------
13 : */
14 :
15 : #include "postgres.h"
16 :
17 : #include "executor/executor.h"
18 : #include "executor/execParallel.h"
19 : #include "executor/nodeGatherMerge.h"
20 : #include "executor/tqueue.h"
21 : #include "lib/binaryheap.h"
22 : #include "miscadmin.h"
23 : #include "optimizer/optimizer.h"
24 :
25 : /*
26 : * When we read tuples from workers, it's a good idea to read several at once
27 : * for efficiency when possible: this minimizes context-switching overhead.
28 : * But reading too many at a time wastes memory without improving performance.
29 : * We'll read up to MAX_TUPLE_STORE tuples (in addition to the first one).
30 : */
31 : #define MAX_TUPLE_STORE 10
32 :
33 : /*
34 : * Pending-tuple array for each worker. This holds additional tuples that
35 : * we were able to fetch from the worker, but can't process yet. In addition,
36 : * this struct holds the "done" flag indicating the worker is known to have
37 : * no more tuples. (We do not use this struct for the leader; we don't keep
38 : * any pending tuples for the leader, and the need_to_scan_locally flag serves
39 : * as its "done" indicator.)
40 : */
41 : typedef struct GMReaderTupleBuffer
42 : {
43 : MinimalTuple *tuple; /* array of length MAX_TUPLE_STORE */
44 : int nTuples; /* number of tuples currently stored */
45 : int readCounter; /* index of next tuple to extract */
46 : bool done; /* true if reader is known exhausted */
47 : } GMReaderTupleBuffer;
48 :
49 : static TupleTableSlot *ExecGatherMerge(PlanState *pstate);
50 : static int32 heap_compare_slots(Datum a, Datum b, void *arg);
51 : static TupleTableSlot *gather_merge_getnext(GatherMergeState *gm_state);
52 : static MinimalTuple gm_readnext_tuple(GatherMergeState *gm_state, int nreader,
53 : bool nowait, bool *done);
54 : static void ExecShutdownGatherMergeWorkers(GatherMergeState *node);
55 : static void gather_merge_setup(GatherMergeState *gm_state);
56 : static void gather_merge_init(GatherMergeState *gm_state);
57 : static void gather_merge_clear_tuples(GatherMergeState *gm_state);
58 : static bool gather_merge_readnext(GatherMergeState *gm_state, int reader,
59 : bool nowait);
60 : static void load_tuple_array(GatherMergeState *gm_state, int reader);
61 :
62 : /* ----------------------------------------------------------------
63 : * ExecInitGather
64 : * ----------------------------------------------------------------
65 : */
66 : GatherMergeState *
67 318 : ExecInitGatherMerge(GatherMerge *node, EState *estate, int eflags)
68 : {
69 : GatherMergeState *gm_state;
70 : Plan *outerNode;
71 : TupleDesc tupDesc;
72 :
73 : /* Gather merge node doesn't have innerPlan node. */
74 : Assert(innerPlan(node) == NULL);
75 :
76 : /*
77 : * create state structure
78 : */
79 318 : gm_state = makeNode(GatherMergeState);
80 318 : gm_state->ps.plan = (Plan *) node;
81 318 : gm_state->ps.state = estate;
82 318 : gm_state->ps.ExecProcNode = ExecGatherMerge;
83 :
84 318 : gm_state->initialized = false;
85 318 : gm_state->gm_initialized = false;
86 318 : gm_state->tuples_needed = -1;
87 :
88 : /*
89 : * Miscellaneous initialization
90 : *
91 : * create expression context for node
92 : */
93 318 : ExecAssignExprContext(estate, &gm_state->ps);
94 :
95 : /*
96 : * GatherMerge doesn't support checking a qual (it's always more efficient
97 : * to do it in the child node).
98 : */
99 : Assert(!node->plan.qual);
100 :
101 : /*
102 : * now initialize outer plan
103 : */
104 318 : outerNode = outerPlan(node);
105 318 : outerPlanState(gm_state) = ExecInitNode(outerNode, estate, eflags);
106 :
107 : /*
108 : * Leader may access ExecProcNode result directly (if
109 : * need_to_scan_locally), or from workers via tuple queue. So we can't
110 : * trivially rely on the slot type being fixed for expressions evaluated
111 : * within this node.
112 : */
113 318 : gm_state->ps.outeropsset = true;
114 318 : gm_state->ps.outeropsfixed = false;
115 :
116 : /*
117 : * Store the tuple descriptor into gather merge state, so we can use it
118 : * while initializing the gather merge slots.
119 : */
120 318 : tupDesc = ExecGetResultType(outerPlanState(gm_state));
121 318 : gm_state->tupDesc = tupDesc;
122 :
123 : /*
124 : * Initialize result type and projection.
125 : */
126 318 : ExecInitResultTypeTL(&gm_state->ps);
127 318 : ExecConditionalAssignProjectionInfo(&gm_state->ps, tupDesc, OUTER_VAR);
128 :
129 : /*
130 : * Without projections result slot type is not trivially known, see
131 : * comment above.
132 : */
133 318 : if (gm_state->ps.ps_ProjInfo == NULL)
134 : {
135 306 : gm_state->ps.resultopsset = true;
136 306 : gm_state->ps.resultopsfixed = false;
137 : }
138 :
139 : /*
140 : * initialize sort-key information
141 : */
142 318 : if (node->numCols)
143 : {
144 : int i;
145 :
146 318 : gm_state->gm_nkeys = node->numCols;
147 318 : gm_state->gm_sortkeys =
148 318 : palloc0(sizeof(SortSupportData) * node->numCols);
149 :
150 738 : for (i = 0; i < node->numCols; i++)
151 : {
152 420 : SortSupport sortKey = gm_state->gm_sortkeys + i;
153 :
154 420 : sortKey->ssup_cxt = CurrentMemoryContext;
155 420 : sortKey->ssup_collation = node->collations[i];
156 420 : sortKey->ssup_nulls_first = node->nullsFirst[i];
157 420 : sortKey->ssup_attno = node->sortColIdx[i];
158 :
159 : /*
160 : * We don't perform abbreviated key conversion here, for the same
161 : * reasons that it isn't used in MergeAppend
162 : */
163 420 : sortKey->abbreviate = false;
164 :
165 420 : PrepareSortSupportFromOrderingOp(node->sortOperators[i], sortKey);
166 : }
167 : }
168 :
169 : /* Now allocate the workspace for gather merge */
170 318 : gather_merge_setup(gm_state);
171 :
172 318 : return gm_state;
173 : }
174 :
175 : /* ----------------------------------------------------------------
176 : * ExecGatherMerge(node)
177 : *
178 : * Scans the relation via multiple workers and returns
179 : * the next qualifying tuple.
180 : * ----------------------------------------------------------------
181 : */
182 : static TupleTableSlot *
183 254142 : ExecGatherMerge(PlanState *pstate)
184 : {
185 254142 : GatherMergeState *node = castNode(GatherMergeState, pstate);
186 : TupleTableSlot *slot;
187 : ExprContext *econtext;
188 :
189 254142 : CHECK_FOR_INTERRUPTS();
190 :
191 : /*
192 : * As with Gather, we don't launch workers until this node is actually
193 : * executed.
194 : */
195 254142 : if (!node->initialized)
196 : {
197 156 : EState *estate = node->ps.state;
198 156 : GatherMerge *gm = castNode(GatherMerge, node->ps.plan);
199 :
200 : /*
201 : * Sometimes we might have to run without parallelism; but if parallel
202 : * mode is active then we can try to fire up some workers.
203 : */
204 156 : if (gm->num_workers > 0 && estate->es_use_parallel_mode)
205 : {
206 : ParallelContext *pcxt;
207 :
208 : /* Initialize, or re-initialize, shared state needed by workers. */
209 156 : if (!node->pei)
210 126 : node->pei = ExecInitParallelPlan(outerPlanState(node),
211 : estate,
212 : gm->initParam,
213 : gm->num_workers,
214 : node->tuples_needed);
215 : else
216 30 : ExecParallelReinitialize(outerPlanState(node),
217 30 : node->pei,
218 : gm->initParam);
219 :
220 : /* Try to launch workers. */
221 156 : pcxt = node->pei->pcxt;
222 156 : LaunchParallelWorkers(pcxt);
223 : /* We save # workers launched for the benefit of EXPLAIN */
224 156 : node->nworkers_launched = pcxt->nworkers_launched;
225 :
226 : /*
227 : * Count number of workers originally wanted and actually
228 : * launched.
229 : */
230 156 : estate->es_parallel_workers_to_launch += pcxt->nworkers_to_launch;
231 156 : estate->es_parallel_workers_launched += pcxt->nworkers_launched;
232 :
233 : /* Set up tuple queue readers to read the results. */
234 156 : if (pcxt->nworkers_launched > 0)
235 : {
236 144 : ExecParallelCreateReaders(node->pei);
237 : /* Make a working array showing the active readers */
238 144 : node->nreaders = pcxt->nworkers_launched;
239 144 : node->reader = (TupleQueueReader **)
240 144 : palloc(node->nreaders * sizeof(TupleQueueReader *));
241 144 : memcpy(node->reader, node->pei->reader,
242 144 : node->nreaders * sizeof(TupleQueueReader *));
243 : }
244 : else
245 : {
246 : /* No workers? Then never mind. */
247 12 : node->nreaders = 0;
248 12 : node->reader = NULL;
249 : }
250 : }
251 :
252 : /* allow leader to participate if enabled or no choice */
253 156 : if (parallel_leader_participation || node->nreaders == 0)
254 150 : node->need_to_scan_locally = true;
255 156 : node->initialized = true;
256 : }
257 :
258 : /*
259 : * Reset per-tuple memory context to free any expression evaluation
260 : * storage allocated in the previous tuple cycle.
261 : */
262 254142 : econtext = node->ps.ps_ExprContext;
263 254142 : ResetExprContext(econtext);
264 :
265 : /*
266 : * Get next tuple, either from one of our workers, or by running the plan
267 : * ourselves.
268 : */
269 254142 : slot = gather_merge_getnext(node);
270 254142 : if (TupIsNull(slot))
271 126 : return NULL;
272 :
273 : /* If no projection is required, we're done. */
274 254016 : if (node->ps.ps_ProjInfo == NULL)
275 254016 : return slot;
276 :
277 : /*
278 : * Form the result tuple using ExecProject(), and return it.
279 : */
280 0 : econtext->ecxt_outertuple = slot;
281 0 : return ExecProject(node->ps.ps_ProjInfo);
282 : }
283 :
284 : /* ----------------------------------------------------------------
285 : * ExecEndGatherMerge
286 : *
287 : * frees any storage allocated through C routines.
288 : * ----------------------------------------------------------------
289 : */
290 : void
291 318 : ExecEndGatherMerge(GatherMergeState *node)
292 : {
293 318 : ExecEndNode(outerPlanState(node)); /* let children clean up first */
294 318 : ExecShutdownGatherMerge(node);
295 318 : }
296 :
297 : /* ----------------------------------------------------------------
298 : * ExecShutdownGatherMerge
299 : *
300 : * Destroy the setup for parallel workers including parallel context.
301 : * ----------------------------------------------------------------
302 : */
303 : void
304 444 : ExecShutdownGatherMerge(GatherMergeState *node)
305 : {
306 444 : ExecShutdownGatherMergeWorkers(node);
307 :
308 : /* Now destroy the parallel context. */
309 444 : if (node->pei != NULL)
310 : {
311 126 : ExecParallelCleanup(node->pei);
312 126 : node->pei = NULL;
313 : }
314 444 : }
315 :
316 : /* ----------------------------------------------------------------
317 : * ExecShutdownGatherMergeWorkers
318 : *
319 : * Stop all the parallel workers.
320 : * ----------------------------------------------------------------
321 : */
322 : static void
323 492 : ExecShutdownGatherMergeWorkers(GatherMergeState *node)
324 : {
325 492 : if (node->pei != NULL)
326 156 : ExecParallelFinish(node->pei);
327 :
328 : /* Flush local copy of reader array */
329 492 : if (node->reader)
330 144 : pfree(node->reader);
331 492 : node->reader = NULL;
332 492 : }
333 :
334 : /* ----------------------------------------------------------------
335 : * ExecReScanGatherMerge
336 : *
337 : * Prepare to re-scan the result of a GatherMerge.
338 : * ----------------------------------------------------------------
339 : */
340 : void
341 48 : ExecReScanGatherMerge(GatherMergeState *node)
342 : {
343 48 : GatherMerge *gm = (GatherMerge *) node->ps.plan;
344 48 : PlanState *outerPlan = outerPlanState(node);
345 :
346 : /* Make sure any existing workers are gracefully shut down */
347 48 : ExecShutdownGatherMergeWorkers(node);
348 :
349 : /* Free any unused tuples, so we don't leak memory across rescans */
350 48 : gather_merge_clear_tuples(node);
351 :
352 : /* Mark node so that shared state will be rebuilt at next call */
353 48 : node->initialized = false;
354 48 : node->gm_initialized = false;
355 :
356 : /*
357 : * Set child node's chgParam to tell it that the next scan might deliver a
358 : * different set of rows within the leader process. (The overall rowset
359 : * shouldn't change, but the leader process's subset might; hence nodes
360 : * between here and the parallel table scan node mustn't optimize on the
361 : * assumption of an unchanging rowset.)
362 : */
363 48 : if (gm->rescan_param >= 0)
364 48 : outerPlan->chgParam = bms_add_member(outerPlan->chgParam,
365 : gm->rescan_param);
366 :
367 : /*
368 : * If chgParam of subnode is not null then plan will be re-scanned by
369 : * first ExecProcNode. Note: because this does nothing if we have a
370 : * rescan_param, it's currently guaranteed that parallel-aware child nodes
371 : * will not see a ReScan call until after they get a ReInitializeDSM call.
372 : * That ordering might not be something to rely on, though. A good rule
373 : * of thumb is that ReInitializeDSM should reset only shared state, ReScan
374 : * should reset only local state, and anything that depends on both of
375 : * those steps being finished must wait until the first ExecProcNode call.
376 : */
377 48 : if (outerPlan->chgParam == NULL)
378 0 : ExecReScan(outerPlan);
379 48 : }
380 :
381 : /*
382 : * Set up the data structures that we'll need for Gather Merge.
383 : *
384 : * We allocate these once on the basis of gm->num_workers, which is an
385 : * upper bound for the number of workers we'll actually have. During
386 : * a rescan, we reset the structures to empty. This approach simplifies
387 : * not leaking memory across rescans.
388 : *
389 : * In the gm_slots[] array, index 0 is for the leader, and indexes 1 to n
390 : * are for workers. The values placed into gm_heap correspond to indexes
391 : * in gm_slots[]. The gm_tuple_buffers[] array, however, is indexed from
392 : * 0 to n-1; it has no entry for the leader.
393 : */
394 : static void
395 318 : gather_merge_setup(GatherMergeState *gm_state)
396 : {
397 318 : GatherMerge *gm = castNode(GatherMerge, gm_state->ps.plan);
398 318 : int nreaders = gm->num_workers;
399 : int i;
400 :
401 : /*
402 : * Allocate gm_slots for the number of workers + one more slot for leader.
403 : * Slot 0 is always for the leader. Leader always calls ExecProcNode() to
404 : * read the tuple, and then stores it directly into its gm_slots entry.
405 : * For other slots, code below will call ExecInitExtraTupleSlot() to
406 : * create a slot for the worker's results. Note that during any single
407 : * scan, we might have fewer than num_workers available workers, in which
408 : * case the extra array entries go unused.
409 : */
410 318 : gm_state->gm_slots = (TupleTableSlot **)
411 318 : palloc0((nreaders + 1) * sizeof(TupleTableSlot *));
412 :
413 : /* Allocate the tuple slot and tuple array for each worker */
414 318 : gm_state->gm_tuple_buffers = (GMReaderTupleBuffer *)
415 318 : palloc0(nreaders * sizeof(GMReaderTupleBuffer));
416 :
417 1194 : for (i = 0; i < nreaders; i++)
418 : {
419 : /* Allocate the tuple array with length MAX_TUPLE_STORE */
420 1752 : gm_state->gm_tuple_buffers[i].tuple =
421 876 : (MinimalTuple *) palloc0(sizeof(MinimalTuple) * MAX_TUPLE_STORE);
422 :
423 : /* Initialize tuple slot for worker */
424 876 : gm_state->gm_slots[i + 1] =
425 876 : ExecInitExtraTupleSlot(gm_state->ps.state, gm_state->tupDesc,
426 : &TTSOpsMinimalTuple);
427 : }
428 :
429 : /* Allocate the resources for the merge */
430 318 : gm_state->gm_heap = binaryheap_allocate(nreaders + 1,
431 : heap_compare_slots,
432 : gm_state);
433 318 : }
434 :
435 : /*
436 : * Initialize the Gather Merge.
437 : *
438 : * Reset data structures to ensure they're empty. Then pull at least one
439 : * tuple from leader + each worker (or set its "done" indicator), and set up
440 : * the heap.
441 : */
442 : static void
443 156 : gather_merge_init(GatherMergeState *gm_state)
444 : {
445 156 : int nreaders = gm_state->nreaders;
446 156 : bool nowait = true;
447 : int i;
448 :
449 : /* Assert that gather_merge_setup made enough space */
450 : Assert(nreaders <= castNode(GatherMerge, gm_state->ps.plan)->num_workers);
451 :
452 : /* Reset leader's tuple slot to empty */
453 156 : gm_state->gm_slots[0] = NULL;
454 :
455 : /* Reset the tuple slot and tuple array for each worker */
456 564 : for (i = 0; i < nreaders; i++)
457 : {
458 : /* Reset tuple array to empty */
459 408 : gm_state->gm_tuple_buffers[i].nTuples = 0;
460 408 : gm_state->gm_tuple_buffers[i].readCounter = 0;
461 : /* Reset done flag to not-done */
462 408 : gm_state->gm_tuple_buffers[i].done = false;
463 : /* Ensure output slot is empty */
464 408 : ExecClearTuple(gm_state->gm_slots[i + 1]);
465 : }
466 :
467 : /* Reset binary heap to empty */
468 156 : binaryheap_reset(gm_state->gm_heap);
469 :
470 : /*
471 : * First, try to read a tuple from each worker (including leader) in
472 : * nowait mode. After this, if not all workers were able to produce a
473 : * tuple (or a "done" indication), then re-read from remaining workers,
474 : * this time using wait mode. Add all live readers (those producing at
475 : * least one tuple) to the heap.
476 : */
477 300 : reread:
478 1416 : for (i = 0; i <= nreaders; i++)
479 : {
480 1116 : CHECK_FOR_INTERRUPTS();
481 :
482 : /* skip this source if already known done */
483 1932 : if ((i == 0) ? gm_state->need_to_scan_locally :
484 816 : !gm_state->gm_tuple_buffers[i - 1].done)
485 : {
486 1094 : if (TupIsNull(gm_state->gm_slots[i]))
487 : {
488 : /* Don't have a tuple yet, try to get one */
489 954 : if (gather_merge_readnext(gm_state, i, nowait))
490 164 : binaryheap_add_unordered(gm_state->gm_heap,
491 : Int32GetDatum(i));
492 : }
493 : else
494 : {
495 : /*
496 : * We already got at least one tuple from this worker, but
497 : * might as well see if it has any more ready by now.
498 : */
499 140 : load_tuple_array(gm_state, i);
500 : }
501 : }
502 : }
503 :
504 : /* need not recheck leader, since nowait doesn't matter for it */
505 710 : for (i = 1; i <= nreaders; i++)
506 : {
507 554 : if (!gm_state->gm_tuple_buffers[i - 1].done &&
508 154 : TupIsNull(gm_state->gm_slots[i]))
509 : {
510 144 : nowait = false;
511 144 : goto reread;
512 : }
513 : }
514 :
515 : /* Now heapify the heap. */
516 156 : binaryheap_build(gm_state->gm_heap);
517 :
518 156 : gm_state->gm_initialized = true;
519 156 : }
520 :
521 : /*
522 : * Clear out the tuple table slot, and any unused pending tuples,
523 : * for each gather merge input.
524 : */
525 : static void
526 174 : gather_merge_clear_tuples(GatherMergeState *gm_state)
527 : {
528 : int i;
529 :
530 648 : for (i = 0; i < gm_state->nreaders; i++)
531 : {
532 474 : GMReaderTupleBuffer *tuple_buffer = &gm_state->gm_tuple_buffers[i];
533 :
534 474 : while (tuple_buffer->readCounter < tuple_buffer->nTuples)
535 0 : pfree(tuple_buffer->tuple[tuple_buffer->readCounter++]);
536 :
537 474 : ExecClearTuple(gm_state->gm_slots[i + 1]);
538 : }
539 174 : }
540 :
541 : /*
542 : * Read the next tuple for gather merge.
543 : *
544 : * Fetch the sorted tuple out of the heap.
545 : */
546 : static TupleTableSlot *
547 254142 : gather_merge_getnext(GatherMergeState *gm_state)
548 : {
549 : int i;
550 :
551 254142 : if (!gm_state->gm_initialized)
552 : {
553 : /*
554 : * First time through: pull the first tuple from each participant, and
555 : * set up the heap.
556 : */
557 156 : gather_merge_init(gm_state);
558 : }
559 : else
560 : {
561 : /*
562 : * Otherwise, pull the next tuple from whichever participant we
563 : * returned from last time, and reinsert that participant's index into
564 : * the heap, because it might now compare differently against the
565 : * other elements of the heap.
566 : */
567 253986 : i = DatumGetInt32(binaryheap_first(gm_state->gm_heap));
568 :
569 253986 : if (gather_merge_readnext(gm_state, i, false))
570 253852 : binaryheap_replace_first(gm_state->gm_heap, Int32GetDatum(i));
571 : else
572 : {
573 : /* reader exhausted, remove it from heap */
574 134 : (void) binaryheap_remove_first(gm_state->gm_heap);
575 : }
576 : }
577 :
578 254142 : if (binaryheap_empty(gm_state->gm_heap))
579 : {
580 : /* All the queues are exhausted, and so is the heap */
581 126 : gather_merge_clear_tuples(gm_state);
582 126 : return NULL;
583 : }
584 : else
585 : {
586 : /* Return next tuple from whichever participant has the leading one */
587 254016 : i = DatumGetInt32(binaryheap_first(gm_state->gm_heap));
588 254016 : return gm_state->gm_slots[i];
589 : }
590 : }
591 :
592 : /*
593 : * Read tuple(s) for given reader in nowait mode, and load into its tuple
594 : * array, until we have MAX_TUPLE_STORE of them or would have to block.
595 : */
596 : static void
597 164 : load_tuple_array(GatherMergeState *gm_state, int reader)
598 : {
599 : GMReaderTupleBuffer *tuple_buffer;
600 : int i;
601 :
602 : /* Don't do anything if this is the leader. */
603 164 : if (reader == 0)
604 138 : return;
605 :
606 26 : tuple_buffer = &gm_state->gm_tuple_buffers[reader - 1];
607 :
608 : /* If there's nothing in the array, reset the counters to zero. */
609 26 : if (tuple_buffer->nTuples == tuple_buffer->readCounter)
610 24 : tuple_buffer->nTuples = tuple_buffer->readCounter = 0;
611 :
612 : /* Try to fill additional slots in the array. */
613 242 : for (i = tuple_buffer->nTuples; i < MAX_TUPLE_STORE; i++)
614 : {
615 : MinimalTuple tuple;
616 :
617 230 : tuple = gm_readnext_tuple(gm_state,
618 : reader,
619 : true,
620 : &tuple_buffer->done);
621 230 : if (!tuple)
622 14 : break;
623 216 : tuple_buffer->tuple[i] = tuple;
624 216 : tuple_buffer->nTuples++;
625 : }
626 : }
627 :
628 : /*
629 : * Store the next tuple for a given reader into the appropriate slot.
630 : *
631 : * Returns true if successful, false if not (either reader is exhausted,
632 : * or we didn't want to wait for a tuple). Sets done flag if reader
633 : * is found to be exhausted.
634 : */
635 : static bool
636 254940 : gather_merge_readnext(GatherMergeState *gm_state, int reader, bool nowait)
637 : {
638 : GMReaderTupleBuffer *tuple_buffer;
639 : MinimalTuple tup;
640 :
641 : /*
642 : * If we're being asked to generate a tuple from the leader, then we just
643 : * call ExecProcNode as normal to produce one.
644 : */
645 254940 : if (reader == 0)
646 : {
647 253896 : if (gm_state->need_to_scan_locally)
648 : {
649 253896 : PlanState *outerPlan = outerPlanState(gm_state);
650 : TupleTableSlot *outerTupleSlot;
651 253896 : EState *estate = gm_state->ps.state;
652 :
653 : /* Install our DSA area while executing the plan. */
654 253896 : estate->es_query_dsa = gm_state->pei ? gm_state->pei->area : NULL;
655 253896 : outerTupleSlot = ExecProcNode(outerPlan);
656 253896 : estate->es_query_dsa = NULL;
657 :
658 253896 : if (!TupIsNull(outerTupleSlot))
659 : {
660 253776 : gm_state->gm_slots[0] = outerTupleSlot;
661 253776 : return true;
662 : }
663 : /* need_to_scan_locally serves as "done" flag for leader */
664 120 : gm_state->need_to_scan_locally = false;
665 : }
666 120 : return false;
667 : }
668 :
669 : /* Otherwise, check the state of the relevant tuple buffer. */
670 1044 : tuple_buffer = &gm_state->gm_tuple_buffers[reader - 1];
671 :
672 1044 : if (tuple_buffer->nTuples > tuple_buffer->readCounter)
673 : {
674 : /* Return any tuple previously read that is still buffered. */
675 216 : tup = tuple_buffer->tuple[tuple_buffer->readCounter++];
676 : }
677 828 : else if (tuple_buffer->done)
678 : {
679 : /* Reader is known to be exhausted. */
680 14 : return false;
681 : }
682 : else
683 : {
684 : /* Read and buffer next tuple. */
685 814 : tup = gm_readnext_tuple(gm_state,
686 : reader,
687 : nowait,
688 : &tuple_buffer->done);
689 814 : if (!tup)
690 790 : return false;
691 :
692 : /*
693 : * Attempt to read more tuples in nowait mode and store them in the
694 : * pending-tuple array for the reader.
695 : */
696 24 : load_tuple_array(gm_state, reader);
697 : }
698 :
699 : Assert(tup);
700 :
701 : /* Build the TupleTableSlot for the given tuple */
702 240 : ExecStoreMinimalTuple(tup, /* tuple to store */
703 240 : gm_state->gm_slots[reader], /* slot in which to
704 : * store the tuple */
705 : true); /* pfree tuple when done with it */
706 :
707 240 : return true;
708 : }
709 :
710 : /*
711 : * Attempt to read a tuple from given worker.
712 : */
713 : static MinimalTuple
714 1044 : gm_readnext_tuple(GatherMergeState *gm_state, int nreader, bool nowait,
715 : bool *done)
716 : {
717 : TupleQueueReader *reader;
718 : MinimalTuple tup;
719 :
720 : /* Check for async events, particularly messages from workers. */
721 1044 : CHECK_FOR_INTERRUPTS();
722 :
723 : /*
724 : * Attempt to read a tuple.
725 : *
726 : * Note that TupleQueueReaderNext will just return NULL for a worker which
727 : * fails to initialize. We'll treat that worker as having produced no
728 : * tuples; WaitForParallelWorkersToFinish will error out when we get
729 : * there.
730 : */
731 1044 : reader = gm_state->reader[nreader - 1];
732 1044 : tup = TupleQueueReaderNext(reader, nowait, done);
733 :
734 : /*
735 : * Since we'll be buffering these across multiple calls, we need to make a
736 : * copy.
737 : */
738 1044 : return tup ? heap_copy_minimal_tuple(tup) : NULL;
739 : }
740 :
741 : /*
742 : * We have one slot for each item in the heap array. We use SlotNumber
743 : * to store slot indexes. This doesn't actually provide any formal
744 : * type-safety, but it makes the code more self-documenting.
745 : */
746 : typedef int32 SlotNumber;
747 :
748 : /*
749 : * Compare the tuples in the two given slots.
750 : */
751 : static int32
752 444 : heap_compare_slots(Datum a, Datum b, void *arg)
753 : {
754 444 : GatherMergeState *node = (GatherMergeState *) arg;
755 444 : SlotNumber slot1 = DatumGetInt32(a);
756 444 : SlotNumber slot2 = DatumGetInt32(b);
757 :
758 444 : TupleTableSlot *s1 = node->gm_slots[slot1];
759 444 : TupleTableSlot *s2 = node->gm_slots[slot2];
760 : int nkey;
761 :
762 : Assert(!TupIsNull(s1));
763 : Assert(!TupIsNull(s2));
764 :
765 672 : for (nkey = 0; nkey < node->gm_nkeys; nkey++)
766 : {
767 444 : SortSupport sortKey = node->gm_sortkeys + nkey;
768 444 : AttrNumber attno = sortKey->ssup_attno;
769 : Datum datum1,
770 : datum2;
771 : bool isNull1,
772 : isNull2;
773 : int compare;
774 :
775 444 : datum1 = slot_getattr(s1, attno, &isNull1);
776 444 : datum2 = slot_getattr(s2, attno, &isNull2);
777 :
778 444 : compare = ApplySortComparator(datum1, isNull1,
779 : datum2, isNull2,
780 : sortKey);
781 444 : if (compare != 0)
782 : {
783 216 : INVERT_COMPARE_RESULT(compare);
784 216 : return compare;
785 : }
786 : }
787 228 : return 0;
788 : }
|
