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 : 170515 : ExecGatherMerge(PlanState *pstate)
185 : : {
186 : 170515 : GatherMergeState *node = castNode(GatherMergeState, pstate);
187 : : TupleTableSlot *slot;
188 : : ExprContext *econtext;
189 : :
190 [ + + ]: 170515 : CHECK_FOR_INTERRUPTS();
191 : :
192 : : /*
193 : : * As with Gather, we don't launch workers until this node is actually
194 : : * executed.
195 : : */
196 [ + + ]: 170515 : 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 = palloc_array(TupleQueueReader *, node->nreaders);
241 : 120 : memcpy(node->reader, node->pei->reader,
242 : 120 : node->nreaders * sizeof(TupleQueueReader *));
243 : : }
244 : : else
245 : : {
246 : : /* No workers? Then never mind. */
247 : 8 : node->nreaders = 0;
248 : 8 : node->reader = NULL;
249 : : }
250 : : }
251 : :
252 : : /* allow leader to participate if enabled or no choice */
253 [ + + + + ]: 128 : if (parallel_leader_participation || node->nreaders == 0)
254 : 124 : node->need_to_scan_locally = true;
255 : 128 : 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 : 170515 : econtext = node->ps.ps_ExprContext;
263 : 170515 : ResetExprContext(econtext);
264 : :
265 : : /*
266 : : * Get next tuple, either from one of our workers, or by running the plan
267 : : * ourselves.
268 : : */
269 : 170515 : slot = gather_merge_getnext(node);
270 [ + + - + ]: 170515 : if (TupIsNull(slot))
271 : 104 : return NULL;
272 : :
273 : : /* If no projection is required, we're done. */
274 [ + - ]: 170411 : if (node->ps.ps_ProjInfo == NULL)
275 : 170411 : 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 : 260 : ExecEndGatherMerge(GatherMergeState *node)
292 : : {
293 : 260 : ExecEndNode(outerPlanState(node)); /* let children clean up first */
294 : 260 : ExecShutdownGatherMerge(node);
295 : 260 : }
296 : :
297 : : /* ----------------------------------------------------------------
298 : : * ExecShutdownGatherMerge
299 : : *
300 : : * Destroy the setup for parallel workers including parallel context.
301 : : * ----------------------------------------------------------------
302 : : */
303 : : void
304 : 368 : ExecShutdownGatherMerge(GatherMergeState *node)
305 : : {
306 : 368 : ExecShutdownGatherMergeWorkers(node);
307 : :
308 : : /* Now destroy the parallel context. */
309 [ + + ]: 368 : if (node->pei != NULL)
310 : : {
311 : 108 : ExecParallelCleanup(node->pei);
312 : 108 : node->pei = NULL;
313 : : }
314 : 368 : }
315 : :
316 : : /* ----------------------------------------------------------------
317 : : * ExecShutdownGatherMergeWorkers
318 : : *
319 : : * Stop all the parallel workers.
320 : : * ----------------------------------------------------------------
321 : : */
322 : : static void
323 : 400 : ExecShutdownGatherMergeWorkers(GatherMergeState *node)
324 : : {
325 [ + + ]: 400 : if (node->pei != NULL)
326 : 128 : ExecParallelFinish(node->pei);
327 : :
328 : : /* Flush local copy of reader array */
329 [ + + ]: 400 : if (node->reader)
330 : 120 : pfree(node->reader);
331 : 400 : node->reader = NULL;
332 : 400 : }
333 : :
334 : : /* ----------------------------------------------------------------
335 : : * ExecReScanGatherMerge
336 : : *
337 : : * Prepare to re-scan the result of a GatherMerge.
338 : : * ----------------------------------------------------------------
339 : : */
340 : : void
341 : 32 : ExecReScanGatherMerge(GatherMergeState *node)
342 : : {
343 : 32 : GatherMerge *gm = (GatherMerge *) node->ps.plan;
344 : 32 : PlanState *outerPlan = outerPlanState(node);
345 : :
346 : : /* Make sure any existing workers are gracefully shut down */
347 : 32 : ExecShutdownGatherMergeWorkers(node);
348 : :
349 : : /* Free any unused tuples, so we don't leak memory across rescans */
350 : 32 : gather_merge_clear_tuples(node);
351 : :
352 : : /* Mark node so that shared state will be rebuilt at next call */
353 : 32 : node->initialized = false;
354 : 32 : 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 [ + - ]: 32 : if (gm->rescan_param >= 0)
364 : 32 : 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 [ - + ]: 32 : if (outerPlan->chgParam == NULL)
378 : 0 : ExecReScan(outerPlan);
379 : 32 : }
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 : 260 : gather_merge_setup(GatherMergeState *gm_state)
396 : : {
397 : 260 : GatherMerge *gm = castNode(GatherMerge, gm_state->ps.plan);
398 : 260 : 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 : 260 : gm_state->gm_slots = palloc0_array(TupleTableSlot *, nreaders + 1);
411 : :
412 : : /* Allocate the tuple slot and tuple array for each worker */
413 : 260 : gm_state->gm_tuple_buffers = palloc0_array(GMReaderTupleBuffer, nreaders);
414 : :
415 [ + + ]: 924 : for (i = 0; i < nreaders; i++)
416 : : {
417 : : /* Allocate the tuple array with length MAX_TUPLE_STORE */
418 : 664 : gm_state->gm_tuple_buffers[i].tuple = palloc0_array(MinimalTuple, MAX_TUPLE_STORE);
419 : :
420 : : /* Initialize tuple slot for worker */
421 : 664 : gm_state->gm_slots[i + 1] =
422 : 664 : ExecInitExtraTupleSlot(gm_state->ps.state, gm_state->tupDesc,
423 : : &TTSOpsMinimalTuple);
424 : : }
425 : :
426 : : /* Allocate the resources for the merge */
427 : 260 : gm_state->gm_heap = binaryheap_allocate(nreaders + 1,
428 : : heap_compare_slots,
429 : : gm_state);
430 : 260 : }
431 : :
432 : : /*
433 : : * Initialize the Gather Merge.
434 : : *
435 : : * Reset data structures to ensure they're empty. Then pull at least one
436 : : * tuple from leader + each worker (or set its "done" indicator), and set up
437 : : * the heap.
438 : : */
439 : : static void
440 : 128 : gather_merge_init(GatherMergeState *gm_state)
441 : : {
442 : 128 : int nreaders = gm_state->nreaders;
443 : 128 : bool nowait = true;
444 : : int i;
445 : :
446 : : /* Assert that gather_merge_setup made enough space */
447 : : Assert(nreaders <= castNode(GatherMerge, gm_state->ps.plan)->num_workers);
448 : :
449 : : /* Reset leader's tuple slot to empty */
450 : 128 : gm_state->gm_slots[0] = NULL;
451 : :
452 : : /* Reset the tuple slot and tuple array for each worker */
453 [ + + ]: 444 : for (i = 0; i < nreaders; i++)
454 : : {
455 : : /* Reset tuple array to empty */
456 : 316 : gm_state->gm_tuple_buffers[i].nTuples = 0;
457 : 316 : gm_state->gm_tuple_buffers[i].readCounter = 0;
458 : : /* Reset done flag to not-done */
459 : 316 : gm_state->gm_tuple_buffers[i].done = false;
460 : : /* Ensure output slot is empty */
461 : 316 : ExecClearTuple(gm_state->gm_slots[i + 1]);
462 : : }
463 : :
464 : : /* Reset binary heap to empty */
465 : 128 : binaryheap_reset(gm_state->gm_heap);
466 : :
467 : : /*
468 : : * First, try to read a tuple from each worker (including leader) in
469 : : * nowait mode. After this, if not all workers were able to produce a
470 : : * tuple (or a "done" indication), then re-read from remaining workers,
471 : : * this time using wait mode. Add all live readers (those producing at
472 : : * least one tuple) to the heap.
473 : : */
474 : 240 : reread:
475 [ + + ]: 1096 : for (i = 0; i <= nreaders; i++)
476 : : {
477 [ + + ]: 856 : CHECK_FOR_INTERRUPTS();
478 : :
479 : : /* skip this source if already known done */
480 [ + + + + ]: 1472 : if ((i == 0) ? gm_state->need_to_scan_locally :
481 : 616 : !gm_state->gm_tuple_buffers[i - 1].done)
482 : : {
483 [ + + + + ]: 823 : if (TupIsNull(gm_state->gm_slots[i]))
484 : : {
485 : : /* Don't have a tuple yet, try to get one */
486 [ + + ]: 922 : if (gather_merge_readnext(gm_state, i, nowait))
487 : 227 : binaryheap_add_unordered(gm_state->gm_heap,
488 : : Int32GetDatum(i));
489 : : }
490 : : else
491 : : {
492 : : /*
493 : : * We already got at least one tuple from this worker, but
494 : : * might as well see if it has any more ready by now.
495 : : */
496 : 128 : load_tuple_array(gm_state, i);
497 : : }
498 : : }
499 : : }
500 : :
501 : : /* need not recheck leader, since nowait doesn't matter for it */
502 [ + + ]: 570 : for (i = 1; i <= nreaders; i++)
503 : : {
504 [ + + ]: 442 : if (!gm_state->gm_tuple_buffers[i - 1].done &&
505 [ + - + + ]: 190 : TupIsNull(gm_state->gm_slots[i]))
506 : : {
507 : 112 : nowait = false;
508 : 112 : goto reread;
509 : : }
510 : : }
511 : :
512 : : /* Now heapify the heap. */
513 : 128 : binaryheap_build(gm_state->gm_heap);
514 : :
515 : 128 : gm_state->gm_initialized = true;
516 : 128 : }
517 : :
518 : : /*
519 : : * Clear out the tuple table slot, and any unused pending tuples,
520 : : * for each gather merge input.
521 : : */
522 : : static void
523 : 136 : gather_merge_clear_tuples(GatherMergeState *gm_state)
524 : : {
525 : : int i;
526 : :
527 [ + + ]: 492 : for (i = 0; i < gm_state->nreaders; i++)
528 : : {
529 : 356 : GMReaderTupleBuffer *tuple_buffer = &gm_state->gm_tuple_buffers[i];
530 : :
531 [ + + ]: 377 : while (tuple_buffer->readCounter < tuple_buffer->nTuples)
532 : 21 : pfree(tuple_buffer->tuple[tuple_buffer->readCounter++]);
533 : :
534 : 356 : ExecClearTuple(gm_state->gm_slots[i + 1]);
535 : : }
536 : 136 : }
537 : :
538 : : /*
539 : : * Read the next tuple for gather merge.
540 : : *
541 : : * Fetch the sorted tuple out of the heap.
542 : : */
543 : : static TupleTableSlot *
544 : 170515 : gather_merge_getnext(GatherMergeState *gm_state)
545 : : {
546 : : int i;
547 : :
548 [ + + ]: 170515 : if (!gm_state->gm_initialized)
549 : : {
550 : : /*
551 : : * First time through: pull the first tuple from each participant, and
552 : : * set up the heap.
553 : : */
554 : 128 : gather_merge_init(gm_state);
555 : : }
556 : : else
557 : : {
558 : : /*
559 : : * Otherwise, pull the next tuple from whichever participant we
560 : : * returned from last time, and reinsert that participant's index into
561 : : * the heap, because it might now compare differently against the
562 : : * other elements of the heap.
563 : : */
564 : 170387 : i = DatumGetInt32(binaryheap_first(gm_state->gm_heap));
565 : :
566 [ + + ]: 170387 : if (gather_merge_readnext(gm_state, i, false))
567 : 170190 : binaryheap_replace_first(gm_state->gm_heap, Int32GetDatum(i));
568 : : else
569 : : {
570 : : /* reader exhausted, remove it from heap */
571 : 197 : (void) binaryheap_remove_first(gm_state->gm_heap);
572 : : }
573 : : }
574 : :
575 [ + + ]: 170515 : if (binaryheap_empty(gm_state->gm_heap))
576 : : {
577 : : /* All the queues are exhausted, and so is the heap */
578 : 104 : gather_merge_clear_tuples(gm_state);
579 : 104 : return NULL;
580 : : }
581 : : else
582 : : {
583 : : /* Return next tuple from whichever participant has the leading one */
584 : 170411 : i = DatumGetInt32(binaryheap_first(gm_state->gm_heap));
585 : 170411 : return gm_state->gm_slots[i];
586 : : }
587 : : }
588 : :
589 : : /*
590 : : * Read tuple(s) for given reader in nowait mode, and load into its tuple
591 : : * array, until we have MAX_TUPLE_STORE of them or would have to block.
592 : : */
593 : : static void
594 : 2077 : load_tuple_array(GatherMergeState *gm_state, int reader)
595 : : {
596 : : GMReaderTupleBuffer *tuple_buffer;
597 : : int i;
598 : :
599 : : /* Don't do anything if this is the leader. */
600 [ + + ]: 2077 : if (reader == 0)
601 : 108 : return;
602 : :
603 : 1969 : tuple_buffer = &gm_state->gm_tuple_buffers[reader - 1];
604 : :
605 : : /* If there's nothing in the array, reset the counters to zero. */
606 [ + + ]: 1969 : if (tuple_buffer->nTuples == tuple_buffer->readCounter)
607 : 1949 : tuple_buffer->nTuples = tuple_buffer->readCounter = 0;
608 : :
609 : : /* Try to fill additional slots in the array. */
610 [ + + ]: 20902 : for (i = tuple_buffer->nTuples; i < MAX_TUPLE_STORE; i++)
611 : : {
612 : : MinimalTuple tuple;
613 : :
614 : 19048 : tuple = gm_readnext_tuple(gm_state,
615 : : reader,
616 : : true,
617 : : &tuple_buffer->done);
618 [ + + ]: 19048 : if (!tuple)
619 : 115 : break;
620 : 18933 : tuple_buffer->tuple[i] = tuple;
621 : 18933 : tuple_buffer->nTuples++;
622 : : }
623 : : }
624 : :
625 : : /*
626 : : * Store the next tuple for a given reader into the appropriate slot.
627 : : *
628 : : * Returns true if successful, false if not (either reader is exhausted,
629 : : * or we didn't want to wait for a tuple). Sets done flag if reader
630 : : * is found to be exhausted.
631 : : */
632 : : static bool
633 : 171082 : gather_merge_readnext(GatherMergeState *gm_state, int reader, bool nowait)
634 : : {
635 : : GMReaderTupleBuffer *tuple_buffer;
636 : : MinimalTuple tup;
637 : :
638 : : /*
639 : : * If we're being asked to generate a tuple from the leader, then we just
640 : : * call ExecProcNode as normal to produce one.
641 : : */
642 [ + + ]: 171082 : if (reader == 0)
643 : : {
644 [ + - ]: 149664 : if (gm_state->need_to_scan_locally)
645 : : {
646 : 149664 : PlanState *outerPlan = outerPlanState(gm_state);
647 : : TupleTableSlot *outerTupleSlot;
648 : 149664 : EState *estate = gm_state->ps.state;
649 : :
650 : : /* Install our DSA area while executing the plan. */
651 [ + - ]: 149664 : estate->es_query_dsa = gm_state->pei ? gm_state->pei->area : NULL;
652 : 149664 : outerTupleSlot = ExecProcNode(outerPlan);
653 : 149664 : estate->es_query_dsa = NULL;
654 : :
655 [ + + + + ]: 149664 : if (!TupIsNull(outerTupleSlot))
656 : : {
657 : 149564 : gm_state->gm_slots[0] = outerTupleSlot;
658 : 149564 : return true;
659 : : }
660 : : /* need_to_scan_locally serves as "done" flag for leader */
661 : 100 : gm_state->need_to_scan_locally = false;
662 : : }
663 : 100 : return false;
664 : : }
665 : :
666 : : /* Otherwise, check the state of the relevant tuple buffer. */
667 : 21418 : tuple_buffer = &gm_state->gm_tuple_buffers[reader - 1];
668 : :
669 [ + + ]: 21418 : if (tuple_buffer->nTuples > tuple_buffer->readCounter)
670 : : {
671 : : /* Return any tuple previously read that is still buffered. */
672 : 18904 : tup = tuple_buffer->tuple[tuple_buffer->readCounter++];
673 : : }
674 [ + + ]: 2514 : else if (tuple_buffer->done)
675 : : {
676 : : /* Reader is known to be exhausted. */
677 : 89 : return false;
678 : : }
679 : : else
680 : : {
681 : : /* Read and buffer next tuple. */
682 : 2425 : tup = gm_readnext_tuple(gm_state,
683 : : reader,
684 : : nowait,
685 : : &tuple_buffer->done);
686 [ + + ]: 2425 : if (!tup)
687 : 476 : return false;
688 : :
689 : : /*
690 : : * Attempt to read more tuples in nowait mode and store them in the
691 : : * pending-tuple array for the reader.
692 : : */
693 : 1949 : load_tuple_array(gm_state, reader);
694 : : }
695 : :
696 : : Assert(tup);
697 : :
698 : : /* Build the TupleTableSlot for the given tuple */
699 : 20853 : ExecStoreMinimalTuple(tup, /* tuple to store */
700 : 20853 : gm_state->gm_slots[reader], /* slot in which to
701 : : * store the tuple */
702 : : true); /* pfree tuple when done with it */
703 : :
704 : 20853 : return true;
705 : : }
706 : :
707 : : /*
708 : : * Attempt to read a tuple from given worker.
709 : : */
710 : : static MinimalTuple
711 : 21473 : gm_readnext_tuple(GatherMergeState *gm_state, int nreader, bool nowait,
712 : : bool *done)
713 : : {
714 : : TupleQueueReader *reader;
715 : : MinimalTuple tup;
716 : :
717 : : /* Check for async events, particularly messages from workers. */
718 [ - + ]: 21473 : CHECK_FOR_INTERRUPTS();
719 : :
720 : : /*
721 : : * Attempt to read a tuple.
722 : : *
723 : : * Note that TupleQueueReaderNext will just return NULL for a worker which
724 : : * fails to initialize. We'll treat that worker as having produced no
725 : : * tuples; WaitForParallelWorkersToFinish will error out when we get
726 : : * there.
727 : : */
728 : 21473 : reader = gm_state->reader[nreader - 1];
729 : 21473 : tup = TupleQueueReaderNext(reader, nowait, done);
730 : :
731 : : /*
732 : : * Since we'll be buffering these across multiple calls, we need to make a
733 : : * copy.
734 : : */
735 [ + + ]: 21473 : return tup ? heap_copy_minimal_tuple(tup, 0) : NULL;
736 : : }
737 : :
738 : : /*
739 : : * We have one slot for each item in the heap array. We use SlotNumber
740 : : * to store slot indexes. This doesn't actually provide any formal
741 : : * type-safety, but it makes the code more self-documenting.
742 : : */
743 : : typedef int32 SlotNumber;
744 : :
745 : : /*
746 : : * Compare the tuples in the two given slots.
747 : : */
748 : : static int32
749 : 100106 : heap_compare_slots(Datum a, Datum b, void *arg)
750 : : {
751 : 100106 : GatherMergeState *node = (GatherMergeState *) arg;
752 : 100106 : SlotNumber slot1 = DatumGetInt32(a);
753 : 100106 : SlotNumber slot2 = DatumGetInt32(b);
754 : :
755 : 100106 : TupleTableSlot *s1 = node->gm_slots[slot1];
756 : 100106 : TupleTableSlot *s2 = node->gm_slots[slot2];
757 : : int nkey;
758 : :
759 : : Assert(!TupIsNull(s1));
760 : : Assert(!TupIsNull(s2));
761 : :
762 [ + + ]: 137516 : for (nkey = 0; nkey < node->gm_nkeys; nkey++)
763 : : {
764 : 100106 : SortSupport sortKey = node->gm_sortkeys + nkey;
765 : 100106 : AttrNumber attno = sortKey->ssup_attno;
766 : : Datum datum1,
767 : : datum2;
768 : : bool isNull1,
769 : : isNull2;
770 : : int compare;
771 : :
772 : 100106 : datum1 = slot_getattr(s1, attno, &isNull1);
773 : 100106 : datum2 = slot_getattr(s2, attno, &isNull2);
774 : :
775 : 100106 : compare = ApplySortComparator(datum1, isNull1,
776 : : datum2, isNull2,
777 : : sortKey);
778 [ + + ]: 100106 : if (compare != 0)
779 : : {
780 [ + + ]: 62696 : INVERT_COMPARE_RESULT(compare);
781 : 62696 : return compare;
782 : : }
783 : : }
784 : 37410 : return 0;
785 : : }
|
