Branch data 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) 1996-2026, 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 "access/htup_details.h"
18 : : #include "executor/executor.h"
19 : : #include "executor/execParallel.h"
20 : : #include "executor/nodeGatherMerge.h"
21 : : #include "executor/tqueue.h"
22 : : #include "lib/binaryheap.h"
23 : : #include "miscadmin.h"
24 : : #include "optimizer/optimizer.h"
25 : : #include "utils/sortsupport.h"
26 : :
27 : : /*
28 : : * When we read tuples from workers, it's a good idea to read several at once
29 : : * for efficiency when possible: this minimizes context-switching overhead.
30 : : * But reading too many at a time wastes memory without improving performance.
31 : : * We'll read up to MAX_TUPLE_STORE tuples (in addition to the first one).
32 : : */
33 : : #define MAX_TUPLE_STORE 10
34 : :
35 : : /*
36 : : * Pending-tuple array for each worker. This holds additional tuples that
37 : : * we were able to fetch from the worker, but can't process yet. In addition,
38 : : * this struct holds the "done" flag indicating the worker is known to have
39 : : * no more tuples. (We do not use this struct for the leader; we don't keep
40 : : * any pending tuples for the leader, and the need_to_scan_locally flag serves
41 : : * as its "done" indicator.)
42 : : */
43 : : typedef struct GMReaderTupleBuffer
44 : : {
45 : : MinimalTuple *tuple; /* array of length MAX_TUPLE_STORE */
46 : : int nTuples; /* number of tuples currently stored */
47 : : int readCounter; /* index of next tuple to extract */
48 : : bool done; /* true if reader is known exhausted */
49 : : } GMReaderTupleBuffer;
50 : :
51 : : static TupleTableSlot *ExecGatherMerge(PlanState *pstate);
52 : : static int32 heap_compare_slots(Datum a, Datum b, void *arg);
53 : : static TupleTableSlot *gather_merge_getnext(GatherMergeState *gm_state);
54 : : static MinimalTuple gm_readnext_tuple(GatherMergeState *gm_state, int nreader,
55 : : bool nowait, bool *done);
56 : : static void ExecShutdownGatherMergeWorkers(GatherMergeState *node);
57 : : static void gather_merge_setup(GatherMergeState *gm_state);
58 : : static void gather_merge_init(GatherMergeState *gm_state);
59 : : static void gather_merge_clear_tuples(GatherMergeState *gm_state);
60 : : static bool gather_merge_readnext(GatherMergeState *gm_state, int reader,
61 : : bool nowait);
62 : : static void load_tuple_array(GatherMergeState *gm_state, int reader);
63 : :
64 : : /* ----------------------------------------------------------------
65 : : * ExecInitGather
66 : : * ----------------------------------------------------------------
67 : : */
68 : : GatherMergeState *
69 : 260 : ExecInitGatherMerge(GatherMerge *node, EState *estate, int eflags)
70 : : {
71 : : GatherMergeState *gm_state;
72 : : Plan *outerNode;
73 : : TupleDesc tupDesc;
74 : :
75 : : /* Gather merge node doesn't have innerPlan node. */
76 : : Assert(innerPlan(node) == NULL);
77 : :
78 : : /*
79 : : * create state structure
80 : : */
81 : 260 : gm_state = makeNode(GatherMergeState);
82 : 260 : gm_state->ps.plan = (Plan *) node;
83 : 260 : gm_state->ps.state = estate;
84 : 260 : gm_state->ps.ExecProcNode = ExecGatherMerge;
85 : :
86 : 260 : gm_state->initialized = false;
87 : 260 : gm_state->gm_initialized = false;
88 : 260 : gm_state->tuples_needed = -1;
89 : :
90 : : /*
91 : : * Miscellaneous initialization
92 : : *
93 : : * create expression context for node
94 : : */
95 : 260 : ExecAssignExprContext(estate, &gm_state->ps);
96 : :
97 : : /*
98 : : * GatherMerge doesn't support checking a qual (it's always more efficient
99 : : * to do it in the child node).
100 : : */
101 : : Assert(!node->plan.qual);
102 : :
103 : : /*
104 : : * now initialize outer plan
105 : : */
106 : 260 : outerNode = outerPlan(node);
107 : 260 : outerPlanState(gm_state) = ExecInitNode(outerNode, estate, eflags);
108 : :
109 : : /*
110 : : * Leader may access ExecProcNode result directly (if
111 : : * need_to_scan_locally), or from workers via tuple queue. So we can't
112 : : * trivially rely on the slot type being fixed for expressions evaluated
113 : : * within this node.
114 : : */
115 : 260 : gm_state->ps.outeropsset = true;
116 : 260 : gm_state->ps.outeropsfixed = false;
117 : :
118 : : /*
119 : : * Store the tuple descriptor into gather merge state, so we can use it
120 : : * while initializing the gather merge slots.
121 : : */
122 : 260 : tupDesc = ExecGetResultType(outerPlanState(gm_state));
123 : 260 : gm_state->tupDesc = tupDesc;
124 : :
125 : : /*
126 : : * Initialize result type and projection.
127 : : */
128 : 260 : ExecInitResultTypeTL(&gm_state->ps);
129 : 260 : ExecConditionalAssignProjectionInfo(&gm_state->ps, tupDesc, OUTER_VAR);
130 : :
131 : : /*
132 : : * Without projections result slot type is not trivially known, see
133 : : * comment above.
134 : : */
135 [ + + ]: 260 : if (gm_state->ps.ps_ProjInfo == NULL)
136 : : {
137 : 252 : gm_state->ps.resultopsset = true;
138 : 252 : gm_state->ps.resultopsfixed = false;
139 : : }
140 : :
141 : : /*
142 : : * initialize sort-key information
143 : : */
144 [ + - ]: 260 : if (node->numCols)
145 : : {
146 : : int i;
147 : :
148 : 260 : gm_state->gm_nkeys = node->numCols;
149 : 260 : gm_state->gm_sortkeys = palloc0_array(SortSupportData, node->numCols);
150 : :
151 [ + + ]: 620 : for (i = 0; i < node->numCols; i++)
152 : : {
153 : 360 : SortSupport sortKey = gm_state->gm_sortkeys + i;
154 : :
155 : 360 : sortKey->ssup_cxt = CurrentMemoryContext;
156 : 360 : sortKey->ssup_collation = node->collations[i];
157 : 360 : sortKey->ssup_nulls_first = node->nullsFirst[i];
158 : 360 : sortKey->ssup_attno = node->sortColIdx[i];
159 : :
160 : : /*
161 : : * We don't perform abbreviated key conversion here, for the same
162 : : * reasons that it isn't used in MergeAppend
163 : : */
164 : 360 : sortKey->abbreviate = false;
165 : :
166 : 360 : PrepareSortSupportFromOrderingOp(node->sortOperators[i], sortKey);
167 : : }
168 : : }
169 : :
170 : : /* Now allocate the workspace for gather merge */
171 : 260 : gather_merge_setup(gm_state);
172 : :
173 : 260 : return gm_state;
174 : : }
175 : :
176 : : /* ----------------------------------------------------------------
177 : : * ExecGatherMerge(node)
178 : : *
179 : : * Scans the relation via multiple workers and returns
180 : : * the next qualifying tuple.
181 : : * ----------------------------------------------------------------
182 : : */
183 : : static TupleTableSlot *
184 : 170292 : ExecGatherMerge(PlanState *pstate)
185 : : {
186 : 170292 : GatherMergeState *node = castNode(GatherMergeState, pstate);
187 : : TupleTableSlot *slot;
188 : : ExprContext *econtext;
189 : :
190 [ + + ]: 170292 : CHECK_FOR_INTERRUPTS();
191 : :
192 : : /*
193 : : * As with Gather, we don't launch workers until this node is actually
194 : : * executed.
195 : : */
196 [ + + ]: 170292 : if (!node->initialized)
197 : : {
198 : 128 : EState *estate = node->ps.state;
199 : 128 : GatherMerge *gm = castNode(GatherMerge, node->ps.plan);
200 : :
201 : : /*
202 : : * Sometimes we might have to run without parallelism; but if parallel
203 : : * mode is active then we can try to fire up some workers.
204 : : */
205 [ + - + - ]: 128 : if (gm->num_workers > 0 && estate->es_use_parallel_mode)
206 : : {
207 : : ParallelContext *pcxt;
208 : :
209 : : /* Initialize, or re-initialize, shared state needed by workers. */
210 [ + + ]: 128 : if (!node->pei)
211 : 108 : node->pei = ExecInitParallelPlan(outerPlanState(node),
212 : : estate,
213 : : gm->initParam,
214 : : gm->num_workers,
215 : : node->tuples_needed);
216 : : else
217 : 20 : ExecParallelReinitialize(outerPlanState(node),
218 : 20 : node->pei,
219 : : gm->initParam);
220 : :
221 : : /* Try to launch workers. */
222 : 128 : pcxt = node->pei->pcxt;
223 : 128 : LaunchParallelWorkers(pcxt);
224 : : /* We save # workers launched for the benefit of EXPLAIN */
225 : 128 : node->nworkers_launched = pcxt->nworkers_launched;
226 : :
227 : : /*
228 : : * Count number of workers originally wanted and actually
229 : : * launched.
230 : : */
231 : 128 : estate->es_parallel_workers_to_launch += pcxt->nworkers_to_launch;
232 : 128 : estate->es_parallel_workers_launched += pcxt->nworkers_launched;
233 : :
234 : : /* Set up tuple queue readers to read the results. */
235 [ + + ]: 128 : if (pcxt->nworkers_launched > 0)
236 : : {
237 : 120 : ExecParallelCreateReaders(node->pei);
238 : : /* Make a working array showing the active readers */
239 : 120 : node->nreaders = pcxt->nworkers_launched;
240 : 120 : node->reader = (TupleQueueReader **)
241 : 120 : palloc(node->nreaders * sizeof(TupleQueueReader *));
242 : 120 : memcpy(node->reader, node->pei->reader,
243 : 120 : node->nreaders * sizeof(TupleQueueReader *));
244 : : }
245 : : else
246 : : {
247 : : /* No workers? Then never mind. */
248 : 8 : node->nreaders = 0;
249 : 8 : node->reader = NULL;
250 : : }
251 : : }
252 : :
253 : : /* allow leader to participate if enabled or no choice */
254 [ + + + + ]: 128 : if (parallel_leader_participation || node->nreaders == 0)
255 : 124 : node->need_to_scan_locally = true;
256 : 128 : node->initialized = true;
257 : : }
258 : :
259 : : /*
260 : : * Reset per-tuple memory context to free any expression evaluation
261 : : * storage allocated in the previous tuple cycle.
262 : : */
263 : 170292 : econtext = node->ps.ps_ExprContext;
264 : 170292 : ResetExprContext(econtext);
265 : :
266 : : /*
267 : : * Get next tuple, either from one of our workers, or by running the plan
268 : : * ourselves.
269 : : */
270 : 170292 : slot = gather_merge_getnext(node);
271 [ + + - + ]: 170292 : if (TupIsNull(slot))
272 : 104 : return NULL;
273 : :
274 : : /* If no projection is required, we're done. */
275 [ + - ]: 170188 : if (node->ps.ps_ProjInfo == NULL)
276 : 170188 : return slot;
277 : :
278 : : /*
279 : : * Form the result tuple using ExecProject(), and return it.
280 : : */
281 : 0 : econtext->ecxt_outertuple = slot;
282 : 0 : return ExecProject(node->ps.ps_ProjInfo);
283 : : }
284 : :
285 : : /* ----------------------------------------------------------------
286 : : * ExecEndGatherMerge
287 : : *
288 : : * frees any storage allocated through C routines.
289 : : * ----------------------------------------------------------------
290 : : */
291 : : void
292 : 260 : ExecEndGatherMerge(GatherMergeState *node)
293 : : {
294 : 260 : ExecEndNode(outerPlanState(node)); /* let children clean up first */
295 : 260 : ExecShutdownGatherMerge(node);
296 : 260 : }
297 : :
298 : : /* ----------------------------------------------------------------
299 : : * ExecShutdownGatherMerge
300 : : *
301 : : * Destroy the setup for parallel workers including parallel context.
302 : : * ----------------------------------------------------------------
303 : : */
304 : : void
305 : 368 : ExecShutdownGatherMerge(GatherMergeState *node)
306 : : {
307 : 368 : ExecShutdownGatherMergeWorkers(node);
308 : :
309 : : /* Now destroy the parallel context. */
310 [ + + ]: 368 : if (node->pei != NULL)
311 : : {
312 : 108 : ExecParallelCleanup(node->pei);
313 : 108 : node->pei = NULL;
314 : : }
315 : 368 : }
316 : :
317 : : /* ----------------------------------------------------------------
318 : : * ExecShutdownGatherMergeWorkers
319 : : *
320 : : * Stop all the parallel workers.
321 : : * ----------------------------------------------------------------
322 : : */
323 : : static void
324 : 400 : ExecShutdownGatherMergeWorkers(GatherMergeState *node)
325 : : {
326 [ + + ]: 400 : if (node->pei != NULL)
327 : 128 : ExecParallelFinish(node->pei);
328 : :
329 : : /* Flush local copy of reader array */
330 [ + + ]: 400 : if (node->reader)
331 : 120 : pfree(node->reader);
332 : 400 : node->reader = NULL;
333 : 400 : }
334 : :
335 : : /* ----------------------------------------------------------------
336 : : * ExecReScanGatherMerge
337 : : *
338 : : * Prepare to re-scan the result of a GatherMerge.
339 : : * ----------------------------------------------------------------
340 : : */
341 : : void
342 : 32 : ExecReScanGatherMerge(GatherMergeState *node)
343 : : {
344 : 32 : GatherMerge *gm = (GatherMerge *) node->ps.plan;
345 : 32 : PlanState *outerPlan = outerPlanState(node);
346 : :
347 : : /* Make sure any existing workers are gracefully shut down */
348 : 32 : ExecShutdownGatherMergeWorkers(node);
349 : :
350 : : /* Free any unused tuples, so we don't leak memory across rescans */
351 : 32 : gather_merge_clear_tuples(node);
352 : :
353 : : /* Mark node so that shared state will be rebuilt at next call */
354 : 32 : node->initialized = false;
355 : 32 : node->gm_initialized = false;
356 : :
357 : : /*
358 : : * Set child node's chgParam to tell it that the next scan might deliver a
359 : : * different set of rows within the leader process. (The overall rowset
360 : : * shouldn't change, but the leader process's subset might; hence nodes
361 : : * between here and the parallel table scan node mustn't optimize on the
362 : : * assumption of an unchanging rowset.)
363 : : */
364 [ + - ]: 32 : if (gm->rescan_param >= 0)
365 : 32 : outerPlan->chgParam = bms_add_member(outerPlan->chgParam,
366 : : gm->rescan_param);
367 : :
368 : : /*
369 : : * If chgParam of subnode is not null then plan will be re-scanned by
370 : : * first ExecProcNode. Note: because this does nothing if we have a
371 : : * rescan_param, it's currently guaranteed that parallel-aware child nodes
372 : : * will not see a ReScan call until after they get a ReInitializeDSM call.
373 : : * That ordering might not be something to rely on, though. A good rule
374 : : * of thumb is that ReInitializeDSM should reset only shared state, ReScan
375 : : * should reset only local state, and anything that depends on both of
376 : : * those steps being finished must wait until the first ExecProcNode call.
377 : : */
378 [ - + ]: 32 : if (outerPlan->chgParam == NULL)
379 : 0 : ExecReScan(outerPlan);
380 : 32 : }
381 : :
382 : : /*
383 : : * Set up the data structures that we'll need for Gather Merge.
384 : : *
385 : : * We allocate these once on the basis of gm->num_workers, which is an
386 : : * upper bound for the number of workers we'll actually have. During
387 : : * a rescan, we reset the structures to empty. This approach simplifies
388 : : * not leaking memory across rescans.
389 : : *
390 : : * In the gm_slots[] array, index 0 is for the leader, and indexes 1 to n
391 : : * are for workers. The values placed into gm_heap correspond to indexes
392 : : * in gm_slots[]. The gm_tuple_buffers[] array, however, is indexed from
393 : : * 0 to n-1; it has no entry for the leader.
394 : : */
395 : : static void
396 : 260 : gather_merge_setup(GatherMergeState *gm_state)
397 : : {
398 : 260 : GatherMerge *gm = castNode(GatherMerge, gm_state->ps.plan);
399 : 260 : int nreaders = gm->num_workers;
400 : : int i;
401 : :
402 : : /*
403 : : * Allocate gm_slots for the number of workers + one more slot for leader.
404 : : * Slot 0 is always for the leader. Leader always calls ExecProcNode() to
405 : : * read the tuple, and then stores it directly into its gm_slots entry.
406 : : * For other slots, code below will call ExecInitExtraTupleSlot() to
407 : : * create a slot for the worker's results. Note that during any single
408 : : * scan, we might have fewer than num_workers available workers, in which
409 : : * case the extra array entries go unused.
410 : : */
411 : 260 : gm_state->gm_slots = (TupleTableSlot **)
412 : 260 : palloc0((nreaders + 1) * sizeof(TupleTableSlot *));
413 : :
414 : : /* Allocate the tuple slot and tuple array for each worker */
415 : 260 : gm_state->gm_tuple_buffers = (GMReaderTupleBuffer *)
416 : 260 : palloc0(nreaders * sizeof(GMReaderTupleBuffer));
417 : :
418 [ + + ]: 924 : for (i = 0; i < nreaders; i++)
419 : : {
420 : : /* Allocate the tuple array with length MAX_TUPLE_STORE */
421 : 664 : gm_state->gm_tuple_buffers[i].tuple = palloc0_array(MinimalTuple, MAX_TUPLE_STORE);
422 : :
423 : : /* Initialize tuple slot for worker */
424 : 664 : gm_state->gm_slots[i + 1] =
425 : 664 : ExecInitExtraTupleSlot(gm_state->ps.state, gm_state->tupDesc,
426 : : &TTSOpsMinimalTuple);
427 : : }
428 : :
429 : : /* Allocate the resources for the merge */
430 : 260 : gm_state->gm_heap = binaryheap_allocate(nreaders + 1,
431 : : heap_compare_slots,
432 : : gm_state);
433 : 260 : }
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 : 128 : gather_merge_init(GatherMergeState *gm_state)
444 : : {
445 : 128 : int nreaders = gm_state->nreaders;
446 : 128 : 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 : 128 : gm_state->gm_slots[0] = NULL;
454 : :
455 : : /* Reset the tuple slot and tuple array for each worker */
456 [ + + ]: 442 : for (i = 0; i < nreaders; i++)
457 : : {
458 : : /* Reset tuple array to empty */
459 : 314 : gm_state->gm_tuple_buffers[i].nTuples = 0;
460 : 314 : gm_state->gm_tuple_buffers[i].readCounter = 0;
461 : : /* Reset done flag to not-done */
462 : 314 : gm_state->gm_tuple_buffers[i].done = false;
463 : : /* Ensure output slot is empty */
464 : 314 : ExecClearTuple(gm_state->gm_slots[i + 1]);
465 : : }
466 : :
467 : : /* Reset binary heap to empty */
468 : 128 : 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 : 237 : reread:
478 [ + + ]: 1080 : for (i = 0; i <= nreaders; i++)
479 : : {
480 [ - + ]: 843 : CHECK_FOR_INTERRUPTS();
481 : :
482 : : /* skip this source if already known done */
483 [ + + + + ]: 1449 : if ((i == 0) ? gm_state->need_to_scan_locally :
484 : 606 : !gm_state->gm_tuple_buffers[i - 1].done)
485 : : {
486 [ + + + + ]: 814 : if (TupIsNull(gm_state->gm_slots[i]))
487 : : {
488 : : /* Don't have a tuple yet, try to get one */
489 [ + + ]: 897 : if (gather_merge_readnext(gm_state, i, nowait))
490 : 197 : 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 : 114 : load_tuple_array(gm_state, i);
500 : : }
501 : : }
502 : : }
503 : :
504 : : /* need not recheck leader, since nowait doesn't matter for it */
505 [ + + ]: 563 : for (i = 1; i <= nreaders; i++)
506 : : {
507 [ + + ]: 435 : if (!gm_state->gm_tuple_buffers[i - 1].done &&
508 [ + - + + ]: 167 : TupIsNull(gm_state->gm_slots[i]))
509 : : {
510 : 109 : nowait = false;
511 : 109 : goto reread;
512 : : }
513 : : }
514 : :
515 : : /* Now heapify the heap. */
516 : 128 : binaryheap_build(gm_state->gm_heap);
517 : :
518 : 128 : gm_state->gm_initialized = true;
519 : 128 : }
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 : 136 : gather_merge_clear_tuples(GatherMergeState *gm_state)
527 : : {
528 : : int i;
529 : :
530 [ + + ]: 490 : for (i = 0; i < gm_state->nreaders; i++)
531 : : {
532 : 354 : GMReaderTupleBuffer *tuple_buffer = &gm_state->gm_tuple_buffers[i];
533 : :
534 [ - + ]: 354 : while (tuple_buffer->readCounter < tuple_buffer->nTuples)
535 : 0 : pfree(tuple_buffer->tuple[tuple_buffer->readCounter++]);
536 : :
537 : 354 : ExecClearTuple(gm_state->gm_slots[i + 1]);
538 : : }
539 : 136 : }
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 : 170292 : gather_merge_getnext(GatherMergeState *gm_state)
548 : : {
549 : : int i;
550 : :
551 [ + + ]: 170292 : 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 : 128 : 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 : 170164 : i = DatumGetInt32(binaryheap_first(gm_state->gm_heap));
568 : :
569 [ + + ]: 170164 : if (gather_merge_readnext(gm_state, i, false))
570 : 169993 : binaryheap_replace_first(gm_state->gm_heap, Int32GetDatum(i));
571 : : else
572 : : {
573 : : /* reader exhausted, remove it from heap */
574 : 171 : (void) binaryheap_remove_first(gm_state->gm_heap);
575 : : }
576 : : }
577 : :
578 [ + + ]: 170292 : if (binaryheap_empty(gm_state->gm_heap))
579 : : {
580 : : /* All the queues are exhausted, and so is the heap */
581 : 104 : gather_merge_clear_tuples(gm_state);
582 : 104 : return NULL;
583 : : }
584 : : else
585 : : {
586 : : /* Return next tuple from whichever participant has the leading one */
587 : 170188 : i = DatumGetInt32(binaryheap_first(gm_state->gm_heap));
588 : 170188 : 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 : 1059 : 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 [ + + ]: 1059 : if (reader == 0)
604 : 105 : return;
605 : :
606 : 954 : tuple_buffer = &gm_state->gm_tuple_buffers[reader - 1];
607 : :
608 : : /* If there's nothing in the array, reset the counters to zero. */
609 [ + + ]: 954 : if (tuple_buffer->nTuples == tuple_buffer->readCounter)
610 : 945 : tuple_buffer->nTuples = tuple_buffer->readCounter = 0;
611 : :
612 : : /* Try to fill additional slots in the array. */
613 [ + + ]: 10031 : for (i = tuple_buffer->nTuples; i < MAX_TUPLE_STORE; i++)
614 : : {
615 : : MinimalTuple tuple;
616 : :
617 : 9145 : tuple = gm_readnext_tuple(gm_state,
618 : : reader,
619 : : true,
620 : : &tuple_buffer->done);
621 [ + + ]: 9145 : if (!tuple)
622 : 68 : break;
623 : 9077 : tuple_buffer->tuple[i] = tuple;
624 : 9077 : 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 : 170864 : 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 [ + + ]: 170864 : if (reader == 0)
646 : : {
647 [ + - ]: 160277 : if (gm_state->need_to_scan_locally)
648 : : {
649 : 160277 : PlanState *outerPlan = outerPlanState(gm_state);
650 : : TupleTableSlot *outerTupleSlot;
651 : 160277 : EState *estate = gm_state->ps.state;
652 : :
653 : : /* Install our DSA area while executing the plan. */
654 [ + - ]: 160277 : estate->es_query_dsa = gm_state->pei ? gm_state->pei->area : NULL;
655 : 160277 : outerTupleSlot = ExecProcNode(outerPlan);
656 : 160277 : estate->es_query_dsa = NULL;
657 : :
658 [ + + + + ]: 160277 : if (!TupIsNull(outerTupleSlot))
659 : : {
660 : 160177 : gm_state->gm_slots[0] = outerTupleSlot;
661 : 160177 : return true;
662 : : }
663 : : /* need_to_scan_locally serves as "done" flag for leader */
664 : 100 : gm_state->need_to_scan_locally = false;
665 : : }
666 : 100 : return false;
667 : : }
668 : :
669 : : /* Otherwise, check the state of the relevant tuple buffer. */
670 : 10587 : tuple_buffer = &gm_state->gm_tuple_buffers[reader - 1];
671 : :
672 [ + + ]: 10587 : if (tuple_buffer->nTuples > tuple_buffer->readCounter)
673 : : {
674 : : /* Return any tuple previously read that is still buffered. */
675 : 9068 : tup = tuple_buffer->tuple[tuple_buffer->readCounter++];
676 : : }
677 [ + + ]: 1519 : else if (tuple_buffer->done)
678 : : {
679 : : /* Reader is known to be exhausted. */
680 : 65 : return false;
681 : : }
682 : : else
683 : : {
684 : : /* Read and buffer next tuple. */
685 : 1454 : tup = gm_readnext_tuple(gm_state,
686 : : reader,
687 : : nowait,
688 : : &tuple_buffer->done);
689 [ + + ]: 1454 : if (!tup)
690 : 509 : 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 : 945 : load_tuple_array(gm_state, reader);
697 : : }
698 : :
699 : : Assert(tup);
700 : :
701 : : /* Build the TupleTableSlot for the given tuple */
702 : 10013 : ExecStoreMinimalTuple(tup, /* tuple to store */
703 : 10013 : gm_state->gm_slots[reader], /* slot in which to
704 : : * store the tuple */
705 : : true); /* pfree tuple when done with it */
706 : :
707 : 10013 : return true;
708 : : }
709 : :
710 : : /*
711 : : * Attempt to read a tuple from given worker.
712 : : */
713 : : static MinimalTuple
714 : 10599 : 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 [ - + ]: 10599 : 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 : 10599 : reader = gm_state->reader[nreader - 1];
732 : 10599 : 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 [ + + ]: 10599 : return tup ? heap_copy_minimal_tuple(tup, 0) : 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 : 72758 : heap_compare_slots(Datum a, Datum b, void *arg)
753 : : {
754 : 72758 : GatherMergeState *node = (GatherMergeState *) arg;
755 : 72758 : SlotNumber slot1 = DatumGetInt32(a);
756 : 72758 : SlotNumber slot2 = DatumGetInt32(b);
757 : :
758 : 72758 : TupleTableSlot *s1 = node->gm_slots[slot1];
759 : 72758 : TupleTableSlot *s2 = node->gm_slots[slot2];
760 : : int nkey;
761 : :
762 : : Assert(!TupIsNull(s1));
763 : : Assert(!TupIsNull(s2));
764 : :
765 [ + + ]: 116635 : for (nkey = 0; nkey < node->gm_nkeys; nkey++)
766 : : {
767 : 72758 : SortSupport sortKey = node->gm_sortkeys + nkey;
768 : 72758 : AttrNumber attno = sortKey->ssup_attno;
769 : : Datum datum1,
770 : : datum2;
771 : : bool isNull1,
772 : : isNull2;
773 : : int compare;
774 : :
775 : 72758 : datum1 = slot_getattr(s1, attno, &isNull1);
776 : 72758 : datum2 = slot_getattr(s2, attno, &isNull2);
777 : :
778 : 72758 : compare = ApplySortComparator(datum1, isNull1,
779 : : datum2, isNull2,
780 : : sortKey);
781 [ + + ]: 72758 : if (compare != 0)
782 : : {
783 [ + + ]: 28881 : INVERT_COMPARE_RESULT(compare);
784 : 28881 : return compare;
785 : : }
786 : : }
787 : 43877 : return 0;
788 : : }
|
