Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * nodeHash.c
4 : * Routines to hash relations for hashjoin
5 : *
6 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/executor/nodeHash.c
12 : *
13 : * See note on parallelism in nodeHashjoin.c.
14 : *
15 : *-------------------------------------------------------------------------
16 : */
17 : /*
18 : * INTERFACE ROUTINES
19 : * MultiExecHash - generate an in-memory hash table of the relation
20 : * ExecInitHash - initialize node and subnodes
21 : * ExecEndHash - shutdown node and subnodes
22 : */
23 :
24 : #include "postgres.h"
25 :
26 : #include <math.h>
27 : #include <limits.h>
28 :
29 : #include "access/htup_details.h"
30 : #include "access/parallel.h"
31 : #include "catalog/pg_statistic.h"
32 : #include "commands/tablespace.h"
33 : #include "executor/executor.h"
34 : #include "executor/hashjoin.h"
35 : #include "executor/nodeHash.h"
36 : #include "executor/nodeHashjoin.h"
37 : #include "miscadmin.h"
38 : #include "port/pg_bitutils.h"
39 : #include "utils/lsyscache.h"
40 : #include "utils/memutils.h"
41 : #include "utils/syscache.h"
42 : #include "utils/wait_event.h"
43 :
44 : static void ExecHashIncreaseNumBatches(HashJoinTable hashtable);
45 : static void ExecHashIncreaseNumBuckets(HashJoinTable hashtable);
46 : static void ExecParallelHashIncreaseNumBatches(HashJoinTable hashtable);
47 : static void ExecParallelHashIncreaseNumBuckets(HashJoinTable hashtable);
48 : static void ExecHashBuildSkewHash(HashState *hashstate,
49 : HashJoinTable hashtable, Hash *node,
50 : int mcvsToUse);
51 : static void ExecHashSkewTableInsert(HashJoinTable hashtable,
52 : TupleTableSlot *slot,
53 : uint32 hashvalue,
54 : int bucketNumber);
55 : static void ExecHashRemoveNextSkewBucket(HashJoinTable hashtable);
56 :
57 : static void *dense_alloc(HashJoinTable hashtable, Size size);
58 : static HashJoinTuple ExecParallelHashTupleAlloc(HashJoinTable hashtable,
59 : size_t size,
60 : dsa_pointer *shared);
61 : static void MultiExecPrivateHash(HashState *node);
62 : static void MultiExecParallelHash(HashState *node);
63 : static inline HashJoinTuple ExecParallelHashFirstTuple(HashJoinTable hashtable,
64 : int bucketno);
65 : static inline HashJoinTuple ExecParallelHashNextTuple(HashJoinTable hashtable,
66 : HashJoinTuple tuple);
67 : static inline void ExecParallelHashPushTuple(dsa_pointer_atomic *head,
68 : HashJoinTuple tuple,
69 : dsa_pointer tuple_shared);
70 : static void ExecParallelHashJoinSetUpBatches(HashJoinTable hashtable, int nbatch);
71 : static void ExecParallelHashEnsureBatchAccessors(HashJoinTable hashtable);
72 : static void ExecParallelHashRepartitionFirst(HashJoinTable hashtable);
73 : static void ExecParallelHashRepartitionRest(HashJoinTable hashtable);
74 : static HashMemoryChunk ExecParallelHashPopChunkQueue(HashJoinTable hashtable,
75 : dsa_pointer *shared);
76 : static bool ExecParallelHashTuplePrealloc(HashJoinTable hashtable,
77 : int batchno,
78 : size_t size);
79 : static void ExecParallelHashMergeCounters(HashJoinTable hashtable);
80 : static void ExecParallelHashCloseBatchAccessors(HashJoinTable hashtable);
81 :
82 :
83 : /* ----------------------------------------------------------------
84 : * ExecHash
85 : *
86 : * stub for pro forma compliance
87 : * ----------------------------------------------------------------
88 : */
89 : static TupleTableSlot *
90 0 : ExecHash(PlanState *pstate)
91 : {
92 0 : elog(ERROR, "Hash node does not support ExecProcNode call convention");
93 : return NULL;
94 : }
95 :
96 : /* ----------------------------------------------------------------
97 : * MultiExecHash
98 : *
99 : * build hash table for hashjoin, doing partitioning if more
100 : * than one batch is required.
101 : * ----------------------------------------------------------------
102 : */
103 : Node *
104 25488 : MultiExecHash(HashState *node)
105 : {
106 : /* must provide our own instrumentation support */
107 25488 : if (node->ps.instrument)
108 334 : InstrStartNode(node->ps.instrument);
109 :
110 25488 : if (node->parallel_state != NULL)
111 414 : MultiExecParallelHash(node);
112 : else
113 25074 : MultiExecPrivateHash(node);
114 :
115 : /* must provide our own instrumentation support */
116 25488 : if (node->ps.instrument)
117 334 : InstrStopNode(node->ps.instrument, node->hashtable->partialTuples);
118 :
119 : /*
120 : * We do not return the hash table directly because it's not a subtype of
121 : * Node, and so would violate the MultiExecProcNode API. Instead, our
122 : * parent Hashjoin node is expected to know how to fish it out of our node
123 : * state. Ugly but not really worth cleaning up, since Hashjoin knows
124 : * quite a bit more about Hash besides that.
125 : */
126 25488 : return NULL;
127 : }
128 :
129 : /* ----------------------------------------------------------------
130 : * MultiExecPrivateHash
131 : *
132 : * parallel-oblivious version, building a backend-private
133 : * hash table and (if necessary) batch files.
134 : * ----------------------------------------------------------------
135 : */
136 : static void
137 25074 : MultiExecPrivateHash(HashState *node)
138 : {
139 : PlanState *outerNode;
140 : HashJoinTable hashtable;
141 : TupleTableSlot *slot;
142 : ExprContext *econtext;
143 :
144 : /*
145 : * get state info from node
146 : */
147 25074 : outerNode = outerPlanState(node);
148 25074 : hashtable = node->hashtable;
149 :
150 : /*
151 : * set expression context
152 : */
153 25074 : econtext = node->ps.ps_ExprContext;
154 :
155 : /*
156 : * Get all tuples from the node below the Hash node and insert into the
157 : * hash table (or temp files).
158 : */
159 : for (;;)
160 8904488 : {
161 : bool isnull;
162 : Datum hashdatum;
163 :
164 8929562 : slot = ExecProcNode(outerNode);
165 8929562 : if (TupIsNull(slot))
166 : break;
167 : /* We have to compute the hash value */
168 8904488 : econtext->ecxt_outertuple = slot;
169 :
170 8904488 : ResetExprContext(econtext);
171 :
172 8904488 : hashdatum = ExecEvalExprSwitchContext(node->hash_expr, econtext,
173 : &isnull);
174 :
175 8904488 : if (!isnull)
176 : {
177 8904386 : uint32 hashvalue = DatumGetUInt32(hashdatum);
178 : int bucketNumber;
179 :
180 8904386 : bucketNumber = ExecHashGetSkewBucket(hashtable, hashvalue);
181 8904386 : if (bucketNumber != INVALID_SKEW_BUCKET_NO)
182 : {
183 : /* It's a skew tuple, so put it into that hash table */
184 588 : ExecHashSkewTableInsert(hashtable, slot, hashvalue,
185 : bucketNumber);
186 588 : hashtable->skewTuples += 1;
187 : }
188 : else
189 : {
190 : /* Not subject to skew optimization, so insert normally */
191 8903798 : ExecHashTableInsert(hashtable, slot, hashvalue);
192 : }
193 8904386 : hashtable->totalTuples += 1;
194 : }
195 : }
196 :
197 : /* resize the hash table if needed (NTUP_PER_BUCKET exceeded) */
198 25074 : if (hashtable->nbuckets != hashtable->nbuckets_optimal)
199 72 : ExecHashIncreaseNumBuckets(hashtable);
200 :
201 : /* Account for the buckets in spaceUsed (reported in EXPLAIN ANALYZE) */
202 25074 : hashtable->spaceUsed += hashtable->nbuckets * sizeof(HashJoinTuple);
203 25074 : if (hashtable->spaceUsed > hashtable->spacePeak)
204 25028 : hashtable->spacePeak = hashtable->spaceUsed;
205 :
206 25074 : hashtable->partialTuples = hashtable->totalTuples;
207 25074 : }
208 :
209 : /* ----------------------------------------------------------------
210 : * MultiExecParallelHash
211 : *
212 : * parallel-aware version, building a shared hash table and
213 : * (if necessary) batch files using the combined effort of
214 : * a set of co-operating backends.
215 : * ----------------------------------------------------------------
216 : */
217 : static void
218 414 : MultiExecParallelHash(HashState *node)
219 : {
220 : ParallelHashJoinState *pstate;
221 : PlanState *outerNode;
222 : HashJoinTable hashtable;
223 : TupleTableSlot *slot;
224 : ExprContext *econtext;
225 : uint32 hashvalue;
226 : Barrier *build_barrier;
227 : int i;
228 :
229 : /*
230 : * get state info from node
231 : */
232 414 : outerNode = outerPlanState(node);
233 414 : hashtable = node->hashtable;
234 :
235 : /*
236 : * set expression context
237 : */
238 414 : econtext = node->ps.ps_ExprContext;
239 :
240 : /*
241 : * Synchronize the parallel hash table build. At this stage we know that
242 : * the shared hash table has been or is being set up by
243 : * ExecHashTableCreate(), but we don't know if our peers have returned
244 : * from there or are here in MultiExecParallelHash(), and if so how far
245 : * through they are. To find out, we check the build_barrier phase then
246 : * and jump to the right step in the build algorithm.
247 : */
248 414 : pstate = hashtable->parallel_state;
249 414 : build_barrier = &pstate->build_barrier;
250 : Assert(BarrierPhase(build_barrier) >= PHJ_BUILD_ALLOCATE);
251 414 : switch (BarrierPhase(build_barrier))
252 : {
253 194 : case PHJ_BUILD_ALLOCATE:
254 :
255 : /*
256 : * Either I just allocated the initial hash table in
257 : * ExecHashTableCreate(), or someone else is doing that. Either
258 : * way, wait for everyone to arrive here so we can proceed.
259 : */
260 194 : BarrierArriveAndWait(build_barrier, WAIT_EVENT_HASH_BUILD_ALLOCATE);
261 : /* Fall through. */
262 :
263 290 : case PHJ_BUILD_HASH_INNER:
264 :
265 : /*
266 : * It's time to begin hashing, or if we just arrived here then
267 : * hashing is already underway, so join in that effort. While
268 : * hashing we have to be prepared to help increase the number of
269 : * batches or buckets at any time, and if we arrived here when
270 : * that was already underway we'll have to help complete that work
271 : * immediately so that it's safe to access batches and buckets
272 : * below.
273 : */
274 290 : if (PHJ_GROW_BATCHES_PHASE(BarrierAttach(&pstate->grow_batches_barrier)) !=
275 : PHJ_GROW_BATCHES_ELECT)
276 0 : ExecParallelHashIncreaseNumBatches(hashtable);
277 290 : if (PHJ_GROW_BUCKETS_PHASE(BarrierAttach(&pstate->grow_buckets_barrier)) !=
278 : PHJ_GROW_BUCKETS_ELECT)
279 0 : ExecParallelHashIncreaseNumBuckets(hashtable);
280 290 : ExecParallelHashEnsureBatchAccessors(hashtable);
281 290 : ExecParallelHashTableSetCurrentBatch(hashtable, 0);
282 : for (;;)
283 2160192 : {
284 : bool isnull;
285 :
286 2160482 : slot = ExecProcNode(outerNode);
287 2160482 : if (TupIsNull(slot))
288 : break;
289 2160192 : econtext->ecxt_outertuple = slot;
290 :
291 2160192 : ResetExprContext(econtext);
292 :
293 2160192 : hashvalue = DatumGetUInt32(ExecEvalExprSwitchContext(node->hash_expr,
294 : econtext,
295 : &isnull));
296 :
297 2160192 : if (!isnull)
298 2160192 : ExecParallelHashTableInsert(hashtable, slot, hashvalue);
299 2160192 : hashtable->partialTuples++;
300 : }
301 :
302 : /*
303 : * Make sure that any tuples we wrote to disk are visible to
304 : * others before anyone tries to load them.
305 : */
306 1296 : for (i = 0; i < hashtable->nbatch; ++i)
307 1006 : sts_end_write(hashtable->batches[i].inner_tuples);
308 :
309 : /*
310 : * Update shared counters. We need an accurate total tuple count
311 : * to control the empty table optimization.
312 : */
313 290 : ExecParallelHashMergeCounters(hashtable);
314 :
315 290 : BarrierDetach(&pstate->grow_buckets_barrier);
316 290 : BarrierDetach(&pstate->grow_batches_barrier);
317 :
318 : /*
319 : * Wait for everyone to finish building and flushing files and
320 : * counters.
321 : */
322 290 : if (BarrierArriveAndWait(build_barrier,
323 : WAIT_EVENT_HASH_BUILD_HASH_INNER))
324 : {
325 : /*
326 : * Elect one backend to disable any further growth. Batches
327 : * are now fixed. While building them we made sure they'd fit
328 : * in our memory budget when we load them back in later (or we
329 : * tried to do that and gave up because we detected extreme
330 : * skew).
331 : */
332 174 : pstate->growth = PHJ_GROWTH_DISABLED;
333 : }
334 : }
335 :
336 : /*
337 : * We're not yet attached to a batch. We all agree on the dimensions and
338 : * number of inner tuples (for the empty table optimization).
339 : */
340 414 : hashtable->curbatch = -1;
341 414 : hashtable->nbuckets = pstate->nbuckets;
342 414 : hashtable->log2_nbuckets = pg_ceil_log2_32(hashtable->nbuckets);
343 414 : hashtable->totalTuples = pstate->total_tuples;
344 :
345 : /*
346 : * Unless we're completely done and the batch state has been freed, make
347 : * sure we have accessors.
348 : */
349 414 : if (BarrierPhase(build_barrier) < PHJ_BUILD_FREE)
350 414 : ExecParallelHashEnsureBatchAccessors(hashtable);
351 :
352 : /*
353 : * The next synchronization point is in ExecHashJoin's HJ_BUILD_HASHTABLE
354 : * case, which will bring the build phase to PHJ_BUILD_RUN (if it isn't
355 : * there already).
356 : */
357 : Assert(BarrierPhase(build_barrier) == PHJ_BUILD_HASH_OUTER ||
358 : BarrierPhase(build_barrier) == PHJ_BUILD_RUN ||
359 : BarrierPhase(build_barrier) == PHJ_BUILD_FREE);
360 414 : }
361 :
362 : /* ----------------------------------------------------------------
363 : * ExecInitHash
364 : *
365 : * Init routine for Hash node
366 : * ----------------------------------------------------------------
367 : */
368 : HashState *
369 35110 : ExecInitHash(Hash *node, EState *estate, int eflags)
370 : {
371 : HashState *hashstate;
372 :
373 : /* check for unsupported flags */
374 : Assert(!(eflags & (EXEC_FLAG_BACKWARD | EXEC_FLAG_MARK)));
375 :
376 : /*
377 : * create state structure
378 : */
379 35110 : hashstate = makeNode(HashState);
380 35110 : hashstate->ps.plan = (Plan *) node;
381 35110 : hashstate->ps.state = estate;
382 35110 : hashstate->ps.ExecProcNode = ExecHash;
383 : /* delay building hashtable until ExecHashTableCreate() in executor run */
384 35110 : hashstate->hashtable = NULL;
385 :
386 : /*
387 : * Miscellaneous initialization
388 : *
389 : * create expression context for node
390 : */
391 35110 : ExecAssignExprContext(estate, &hashstate->ps);
392 :
393 : /*
394 : * initialize child nodes
395 : */
396 35110 : outerPlanState(hashstate) = ExecInitNode(outerPlan(node), estate, eflags);
397 :
398 : /*
399 : * initialize our result slot and type. No need to build projection
400 : * because this node doesn't do projections.
401 : */
402 35110 : ExecInitResultTupleSlotTL(&hashstate->ps, &TTSOpsMinimalTuple);
403 35110 : hashstate->ps.ps_ProjInfo = NULL;
404 :
405 : Assert(node->plan.qual == NIL);
406 :
407 : /*
408 : * Delay initialization of hash_expr until ExecInitHashJoin(). We cannot
409 : * build the ExprState here as we don't yet know the join type we're going
410 : * to be hashing values for and we need to know that before calling
411 : * ExecBuildHash32Expr as the keep_nulls parameter depends on the join
412 : * type.
413 : */
414 35110 : hashstate->hash_expr = NULL;
415 :
416 35110 : return hashstate;
417 : }
418 :
419 : /* ---------------------------------------------------------------
420 : * ExecEndHash
421 : *
422 : * clean up routine for Hash node
423 : * ----------------------------------------------------------------
424 : */
425 : void
426 34998 : ExecEndHash(HashState *node)
427 : {
428 : PlanState *outerPlan;
429 :
430 : /*
431 : * shut down the subplan
432 : */
433 34998 : outerPlan = outerPlanState(node);
434 34998 : ExecEndNode(outerPlan);
435 34998 : }
436 :
437 :
438 : /* ----------------------------------------------------------------
439 : * ExecHashTableCreate
440 : *
441 : * create an empty hashtable data structure for hashjoin.
442 : * ----------------------------------------------------------------
443 : */
444 : HashJoinTable
445 25488 : ExecHashTableCreate(HashState *state)
446 : {
447 : Hash *node;
448 : HashJoinTable hashtable;
449 : Plan *outerNode;
450 : size_t space_allowed;
451 : int nbuckets;
452 : int nbatch;
453 : double rows;
454 : int num_skew_mcvs;
455 : int log2_nbuckets;
456 : MemoryContext oldcxt;
457 :
458 : /*
459 : * Get information about the size of the relation to be hashed (it's the
460 : * "outer" subtree of this node, but the inner relation of the hashjoin).
461 : * Compute the appropriate size of the hash table.
462 : */
463 25488 : node = (Hash *) state->ps.plan;
464 25488 : outerNode = outerPlan(node);
465 :
466 : /*
467 : * If this is shared hash table with a partial plan, then we can't use
468 : * outerNode->plan_rows to estimate its size. We need an estimate of the
469 : * total number of rows across all copies of the partial plan.
470 : */
471 25488 : rows = node->plan.parallel_aware ? node->rows_total : outerNode->plan_rows;
472 :
473 25074 : ExecChooseHashTableSize(rows, outerNode->plan_width,
474 25488 : OidIsValid(node->skewTable),
475 25488 : state->parallel_state != NULL,
476 25488 : state->parallel_state != NULL ?
477 414 : state->parallel_state->nparticipants - 1 : 0,
478 : &space_allowed,
479 : &nbuckets, &nbatch, &num_skew_mcvs);
480 :
481 : /* nbuckets must be a power of 2 */
482 25488 : log2_nbuckets = pg_ceil_log2_32(nbuckets);
483 : Assert(nbuckets == (1 << log2_nbuckets));
484 :
485 : /*
486 : * Initialize the hash table control block.
487 : *
488 : * The hashtable control block is just palloc'd from the executor's
489 : * per-query memory context. Everything else should be kept inside the
490 : * subsidiary hashCxt, batchCxt or spillCxt.
491 : */
492 25488 : hashtable = palloc_object(HashJoinTableData);
493 25488 : hashtable->nbuckets = nbuckets;
494 25488 : hashtable->nbuckets_original = nbuckets;
495 25488 : hashtable->nbuckets_optimal = nbuckets;
496 25488 : hashtable->log2_nbuckets = log2_nbuckets;
497 25488 : hashtable->log2_nbuckets_optimal = log2_nbuckets;
498 25488 : hashtable->buckets.unshared = NULL;
499 25488 : hashtable->skewEnabled = false;
500 25488 : hashtable->skewBucket = NULL;
501 25488 : hashtable->skewBucketLen = 0;
502 25488 : hashtable->nSkewBuckets = 0;
503 25488 : hashtable->skewBucketNums = NULL;
504 25488 : hashtable->nbatch = nbatch;
505 25488 : hashtable->curbatch = 0;
506 25488 : hashtable->nbatch_original = nbatch;
507 25488 : hashtable->nbatch_outstart = nbatch;
508 25488 : hashtable->growEnabled = true;
509 25488 : hashtable->totalTuples = 0;
510 25488 : hashtable->partialTuples = 0;
511 25488 : hashtable->skewTuples = 0;
512 25488 : hashtable->innerBatchFile = NULL;
513 25488 : hashtable->outerBatchFile = NULL;
514 25488 : hashtable->spaceUsed = 0;
515 25488 : hashtable->spacePeak = 0;
516 25488 : hashtable->spaceAllowed = space_allowed;
517 25488 : hashtable->spaceUsedSkew = 0;
518 25488 : hashtable->spaceAllowedSkew =
519 25488 : hashtable->spaceAllowed * SKEW_HASH_MEM_PERCENT / 100;
520 25488 : hashtable->chunks = NULL;
521 25488 : hashtable->current_chunk = NULL;
522 25488 : hashtable->parallel_state = state->parallel_state;
523 25488 : hashtable->area = state->ps.state->es_query_dsa;
524 25488 : hashtable->batches = NULL;
525 :
526 : #ifdef HJDEBUG
527 : printf("Hashjoin %p: initial nbatch = %d, nbuckets = %d\n",
528 : hashtable, nbatch, nbuckets);
529 : #endif
530 :
531 : /*
532 : * Create temporary memory contexts in which to keep the hashtable working
533 : * storage. See notes in executor/hashjoin.h.
534 : */
535 25488 : hashtable->hashCxt = AllocSetContextCreate(CurrentMemoryContext,
536 : "HashTableContext",
537 : ALLOCSET_DEFAULT_SIZES);
538 :
539 25488 : hashtable->batchCxt = AllocSetContextCreate(hashtable->hashCxt,
540 : "HashBatchContext",
541 : ALLOCSET_DEFAULT_SIZES);
542 :
543 25488 : hashtable->spillCxt = AllocSetContextCreate(hashtable->hashCxt,
544 : "HashSpillContext",
545 : ALLOCSET_DEFAULT_SIZES);
546 :
547 : /* Allocate data that will live for the life of the hashjoin */
548 :
549 25488 : oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
550 :
551 25488 : if (nbatch > 1 && hashtable->parallel_state == NULL)
552 : {
553 : MemoryContext oldctx;
554 :
555 : /*
556 : * allocate and initialize the file arrays in hashCxt (not needed for
557 : * parallel case which uses shared tuplestores instead of raw files)
558 : */
559 126 : oldctx = MemoryContextSwitchTo(hashtable->spillCxt);
560 :
561 126 : hashtable->innerBatchFile = palloc0_array(BufFile *, nbatch);
562 126 : hashtable->outerBatchFile = palloc0_array(BufFile *, nbatch);
563 :
564 126 : MemoryContextSwitchTo(oldctx);
565 :
566 : /* The files will not be opened until needed... */
567 : /* ... but make sure we have temp tablespaces established for them */
568 126 : PrepareTempTablespaces();
569 : }
570 :
571 25488 : MemoryContextSwitchTo(oldcxt);
572 :
573 25488 : if (hashtable->parallel_state)
574 : {
575 414 : ParallelHashJoinState *pstate = hashtable->parallel_state;
576 : Barrier *build_barrier;
577 :
578 : /*
579 : * Attach to the build barrier. The corresponding detach operation is
580 : * in ExecHashTableDetach. Note that we won't attach to the
581 : * batch_barrier for batch 0 yet. We'll attach later and start it out
582 : * in PHJ_BATCH_PROBE phase, because batch 0 is allocated up front and
583 : * then loaded while hashing (the standard hybrid hash join
584 : * algorithm), and we'll coordinate that using build_barrier.
585 : */
586 414 : build_barrier = &pstate->build_barrier;
587 414 : BarrierAttach(build_barrier);
588 :
589 : /*
590 : * So far we have no idea whether there are any other participants,
591 : * and if so, what phase they are working on. The only thing we care
592 : * about at this point is whether someone has already created the
593 : * SharedHashJoinBatch objects and the hash table for batch 0. One
594 : * backend will be elected to do that now if necessary.
595 : */
596 592 : if (BarrierPhase(build_barrier) == PHJ_BUILD_ELECT &&
597 178 : BarrierArriveAndWait(build_barrier, WAIT_EVENT_HASH_BUILD_ELECT))
598 : {
599 174 : pstate->nbatch = nbatch;
600 174 : pstate->space_allowed = space_allowed;
601 174 : pstate->growth = PHJ_GROWTH_OK;
602 :
603 : /* Set up the shared state for coordinating batches. */
604 174 : ExecParallelHashJoinSetUpBatches(hashtable, nbatch);
605 :
606 : /*
607 : * Allocate batch 0's hash table up front so we can load it
608 : * directly while hashing.
609 : */
610 174 : pstate->nbuckets = nbuckets;
611 174 : ExecParallelHashTableAlloc(hashtable, 0);
612 : }
613 :
614 : /*
615 : * The next Parallel Hash synchronization point is in
616 : * MultiExecParallelHash(), which will progress it all the way to
617 : * PHJ_BUILD_RUN. The caller must not return control from this
618 : * executor node between now and then.
619 : */
620 : }
621 : else
622 : {
623 : /*
624 : * Prepare context for the first-scan space allocations; allocate the
625 : * hashbucket array therein, and set each bucket "empty".
626 : */
627 25074 : MemoryContextSwitchTo(hashtable->batchCxt);
628 :
629 25074 : hashtable->buckets.unshared = palloc0_array(HashJoinTuple, nbuckets);
630 :
631 : /*
632 : * Set up for skew optimization, if possible and there's a need for
633 : * more than one batch. (In a one-batch join, there's no point in
634 : * it.)
635 : */
636 25074 : if (nbatch > 1)
637 126 : ExecHashBuildSkewHash(state, hashtable, node, num_skew_mcvs);
638 :
639 25074 : MemoryContextSwitchTo(oldcxt);
640 : }
641 :
642 25488 : return hashtable;
643 : }
644 :
645 :
646 : /*
647 : * Compute appropriate size for hashtable given the estimated size of the
648 : * relation to be hashed (number of rows and average row width).
649 : *
650 : * This is exported so that the planner's costsize.c can use it.
651 : */
652 :
653 : /* Target bucket loading (tuples per bucket) */
654 : #define NTUP_PER_BUCKET 1
655 :
656 : void
657 898818 : ExecChooseHashTableSize(double ntuples, int tupwidth, bool useskew,
658 : bool try_combined_hash_mem,
659 : int parallel_workers,
660 : size_t *space_allowed,
661 : int *numbuckets,
662 : int *numbatches,
663 : int *num_skew_mcvs)
664 : {
665 : int tupsize;
666 : double inner_rel_bytes;
667 : size_t hash_table_bytes;
668 : size_t bucket_bytes;
669 : size_t max_pointers;
670 898818 : int nbatch = 1;
671 : int nbuckets;
672 : double dbuckets;
673 :
674 : /* Force a plausible relation size if no info */
675 898818 : if (ntuples <= 0.0)
676 150 : ntuples = 1000.0;
677 :
678 : /*
679 : * Estimate tupsize based on footprint of tuple in hashtable... note this
680 : * does not allow for any palloc overhead. The manipulations of spaceUsed
681 : * don't count palloc overhead either.
682 : */
683 898818 : tupsize = HJTUPLE_OVERHEAD +
684 898818 : MAXALIGN(SizeofMinimalTupleHeader) +
685 898818 : MAXALIGN(tupwidth);
686 898818 : inner_rel_bytes = ntuples * tupsize;
687 :
688 : /*
689 : * Compute in-memory hashtable size limit from GUCs.
690 : */
691 898818 : hash_table_bytes = get_hash_memory_limit();
692 :
693 : /*
694 : * Parallel Hash tries to use the combined hash_mem of all workers to
695 : * avoid the need to batch. If that won't work, it falls back to hash_mem
696 : * per worker and tries to process batches in parallel.
697 : */
698 898818 : if (try_combined_hash_mem)
699 : {
700 : /* Careful, this could overflow size_t */
701 : double newlimit;
702 :
703 75570 : newlimit = (double) hash_table_bytes * (double) (parallel_workers + 1);
704 75570 : newlimit = Min(newlimit, (double) SIZE_MAX);
705 75570 : hash_table_bytes = (size_t) newlimit;
706 : }
707 :
708 898818 : *space_allowed = hash_table_bytes;
709 :
710 : /*
711 : * If skew optimization is possible, estimate the number of skew buckets
712 : * that will fit in the memory allowed, and decrement the assumed space
713 : * available for the main hash table accordingly.
714 : *
715 : * We make the optimistic assumption that each skew bucket will contain
716 : * one inner-relation tuple. If that turns out to be low, we will recover
717 : * at runtime by reducing the number of skew buckets.
718 : *
719 : * hashtable->skewBucket will have up to 8 times as many HashSkewBucket
720 : * pointers as the number of MCVs we allow, since ExecHashBuildSkewHash
721 : * will round up to the next power of 2 and then multiply by 4 to reduce
722 : * collisions.
723 : */
724 898818 : if (useskew)
725 : {
726 : size_t bytes_per_mcv;
727 : size_t skew_mcvs;
728 :
729 : /*----------
730 : * Compute number of MCVs we could hold in hash_table_bytes
731 : *
732 : * Divisor is:
733 : * size of a hash tuple +
734 : * worst-case size of skewBucket[] per MCV +
735 : * size of skewBucketNums[] entry +
736 : * size of skew bucket struct itself
737 : *----------
738 : */
739 893050 : bytes_per_mcv = tupsize +
740 : (8 * sizeof(HashSkewBucket *)) +
741 893050 : sizeof(int) +
742 : SKEW_BUCKET_OVERHEAD;
743 893050 : skew_mcvs = hash_table_bytes / bytes_per_mcv;
744 :
745 : /*
746 : * Now scale by SKEW_HASH_MEM_PERCENT (we do it in this order so as
747 : * not to worry about size_t overflow in the multiplication)
748 : */
749 893050 : skew_mcvs = (skew_mcvs * SKEW_HASH_MEM_PERCENT) / 100;
750 :
751 : /* Now clamp to integer range */
752 893050 : skew_mcvs = Min(skew_mcvs, INT_MAX);
753 :
754 893050 : *num_skew_mcvs = (int) skew_mcvs;
755 :
756 : /* Reduce hash_table_bytes by the amount needed for the skew table */
757 893050 : if (skew_mcvs > 0)
758 893050 : hash_table_bytes -= skew_mcvs * bytes_per_mcv;
759 : }
760 : else
761 5768 : *num_skew_mcvs = 0;
762 :
763 : /*
764 : * Set nbuckets to achieve an average bucket load of NTUP_PER_BUCKET when
765 : * memory is filled, assuming a single batch; but limit the value so that
766 : * the pointer arrays we'll try to allocate do not exceed hash_table_bytes
767 : * nor MaxAllocSize.
768 : *
769 : * Note that both nbuckets and nbatch must be powers of 2 to make
770 : * ExecHashGetBucketAndBatch fast.
771 : */
772 898818 : max_pointers = hash_table_bytes / sizeof(HashJoinTuple);
773 898818 : max_pointers = Min(max_pointers, MaxAllocSize / sizeof(HashJoinTuple));
774 : /* If max_pointers isn't a power of 2, must round it down to one */
775 898818 : max_pointers = pg_prevpower2_size_t(max_pointers);
776 :
777 : /* Also ensure we avoid integer overflow in nbatch and nbuckets */
778 : /* (this step is redundant given the current value of MaxAllocSize) */
779 898818 : max_pointers = Min(max_pointers, INT_MAX / 2 + 1);
780 :
781 898818 : dbuckets = ceil(ntuples / NTUP_PER_BUCKET);
782 898818 : dbuckets = Min(dbuckets, max_pointers);
783 898818 : nbuckets = (int) dbuckets;
784 : /* don't let nbuckets be really small, though ... */
785 898818 : nbuckets = Max(nbuckets, 1024);
786 : /* ... and force it to be a power of 2. */
787 898818 : nbuckets = pg_nextpower2_32(nbuckets);
788 :
789 : /*
790 : * If there's not enough space to store the projected number of tuples and
791 : * the required bucket headers, we will need multiple batches.
792 : */
793 898818 : bucket_bytes = sizeof(HashJoinTuple) * nbuckets;
794 898818 : if (inner_rel_bytes + bucket_bytes > hash_table_bytes)
795 : {
796 : /* We'll need multiple batches */
797 : size_t sbuckets;
798 : double dbatch;
799 : int minbatch;
800 : size_t bucket_size;
801 :
802 : /*
803 : * If Parallel Hash with combined hash_mem would still need multiple
804 : * batches, we'll have to fall back to regular hash_mem budget.
805 : */
806 5166 : if (try_combined_hash_mem)
807 : {
808 246 : ExecChooseHashTableSize(ntuples, tupwidth, useskew,
809 : false, parallel_workers,
810 : space_allowed,
811 : numbuckets,
812 : numbatches,
813 : num_skew_mcvs);
814 246 : return;
815 : }
816 :
817 : /*
818 : * Estimate the number of buckets we'll want to have when hash_mem is
819 : * entirely full. Each bucket will contain a bucket pointer plus
820 : * NTUP_PER_BUCKET tuples, whose projected size already includes
821 : * overhead for the hash code, pointer to the next tuple, etc.
822 : */
823 4920 : bucket_size = (tupsize * NTUP_PER_BUCKET + sizeof(HashJoinTuple));
824 4920 : if (hash_table_bytes <= bucket_size)
825 0 : sbuckets = 1; /* avoid pg_nextpower2_size_t(0) */
826 : else
827 4920 : sbuckets = pg_nextpower2_size_t(hash_table_bytes / bucket_size);
828 4920 : sbuckets = Min(sbuckets, max_pointers);
829 4920 : nbuckets = (int) sbuckets;
830 4920 : nbuckets = pg_nextpower2_32(nbuckets);
831 4920 : bucket_bytes = nbuckets * sizeof(HashJoinTuple);
832 :
833 : /*
834 : * Buckets are simple pointers to hashjoin tuples, while tupsize
835 : * includes the pointer, hash code, and MinimalTupleData. So buckets
836 : * should never really exceed 25% of hash_mem (even for
837 : * NTUP_PER_BUCKET=1); except maybe for hash_mem values that are not
838 : * 2^N bytes, where we might get more because of doubling. So let's
839 : * look for 50% here.
840 : */
841 : Assert(bucket_bytes <= hash_table_bytes / 2);
842 :
843 : /* Calculate required number of batches. */
844 4920 : dbatch = ceil(inner_rel_bytes / (hash_table_bytes - bucket_bytes));
845 4920 : dbatch = Min(dbatch, max_pointers);
846 4920 : minbatch = (int) dbatch;
847 4920 : nbatch = pg_nextpower2_32(Max(2, minbatch));
848 : }
849 :
850 : /*
851 : * Optimize the total amount of memory consumed by the hash node.
852 : *
853 : * The nbatch calculation above focuses on the in-memory hash table,
854 : * assuming no per-batch overhead. But each batch may have two files, each
855 : * with a BLCKSZ buffer. For large nbatch values these buffers may use
856 : * significantly more memory than the hash table.
857 : *
858 : * The total memory usage may be expressed by this formula:
859 : *
860 : * (inner_rel_bytes / nbatch) + (2 * nbatch * BLCKSZ)
861 : *
862 : * where (inner_rel_bytes / nbatch) is the size of the in-memory hash
863 : * table and (2 * nbatch * BLCKSZ) is the amount of memory used by file
864 : * buffers.
865 : *
866 : * The nbatch calculation however ignores the second part. And for very
867 : * large inner_rel_bytes, there may be no nbatch that keeps total memory
868 : * usage under the budget (work_mem * hash_mem_multiplier). To deal with
869 : * that, we will adjust nbatch to minimize total memory consumption across
870 : * both the hashtable and file buffers.
871 : *
872 : * As we increase the size of the hashtable, the number of batches
873 : * decreases, and the total memory usage follows a U-shaped curve. We find
874 : * the minimum nbatch by "walking back" -- checking if halving nbatch
875 : * would lower the total memory usage. We stop when it no longer helps.
876 : *
877 : * We only reduce the number of batches. Adding batches reduces memory
878 : * usage only when most of the memory is used by the hash table, with
879 : * total memory usage within the limit or not far from it. We don't want
880 : * to start batching when not needed, even if that would reduce memory
881 : * usage.
882 : *
883 : * While growing the hashtable, we also adjust the number of buckets to
884 : * maintain a load factor of NTUP_PER_BUCKET while squeezing tuples back
885 : * from batches into the hashtable.
886 : *
887 : * Note that we can only change nbuckets during initial hashtable sizing.
888 : * Once we start building the hash, nbuckets is fixed (we may still grow
889 : * the hash table).
890 : *
891 : * We double several parameters (space_allowed, nbuckets, num_skew_mcvs),
892 : * which introduces a risk of overflow. We avoid this by exiting the loop.
893 : * We could do something smarter (e.g. capping nbuckets and continue), but
894 : * the complexity is not worth it. Such cases are extremely rare, and this
895 : * is a best-effort attempt to reduce memory usage.
896 : */
897 899406 : while (nbatch > 1)
898 : {
899 : /* Check that buckets wont't overflow MaxAllocSize */
900 5754 : if (nbuckets > (MaxAllocSize / sizeof(HashJoinTuple) / 2))
901 0 : break;
902 :
903 : /* num_skew_mcvs should be less than nbuckets */
904 : Assert((*num_skew_mcvs) < (INT_MAX / 2));
905 :
906 : /*
907 : * Check that space_allowed won't overlow SIZE_MAX.
908 : *
909 : * We don't use hash_table_bytes here, because it does not include the
910 : * skew buckets. And we want to limit the overall memory limit.
911 : */
912 5754 : if ((*space_allowed) > (SIZE_MAX / 2))
913 0 : break;
914 :
915 : /*
916 : * Will halving the number of batches and doubling the size of the
917 : * hashtable reduce overall memory usage?
918 : *
919 : * This is the same as (S = space_allowed):
920 : *
921 : * (S + 2 * nbatch * BLCKSZ) < (S * 2 + nbatch * BLCKSZ)
922 : *
923 : * but avoiding intermediate overflow.
924 : */
925 5754 : if (nbatch < (*space_allowed) / BLCKSZ)
926 4920 : break;
927 :
928 : /*
929 : * MaxAllocSize is sufficiently small that we are not worried about
930 : * overflowing nbuckets.
931 : */
932 834 : nbuckets *= 2;
933 :
934 834 : *num_skew_mcvs = (*num_skew_mcvs) * 2;
935 834 : *space_allowed = (*space_allowed) * 2;
936 :
937 834 : nbatch /= 2;
938 : }
939 :
940 : Assert(nbuckets > 0);
941 : Assert(nbatch > 0);
942 :
943 898572 : *numbuckets = nbuckets;
944 898572 : *numbatches = nbatch;
945 : }
946 :
947 :
948 : /* ----------------------------------------------------------------
949 : * ExecHashTableDestroy
950 : *
951 : * destroy a hash table
952 : * ----------------------------------------------------------------
953 : */
954 : void
955 25378 : ExecHashTableDestroy(HashJoinTable hashtable)
956 : {
957 : int i;
958 :
959 : /*
960 : * Make sure all the temp files are closed. We skip batch 0, since it
961 : * can't have any temp files (and the arrays might not even exist if
962 : * nbatch is only 1). Parallel hash joins don't use these files.
963 : */
964 25378 : if (hashtable->innerBatchFile != NULL)
965 : {
966 1172 : for (i = 1; i < hashtable->nbatch; i++)
967 : {
968 952 : if (hashtable->innerBatchFile[i])
969 0 : BufFileClose(hashtable->innerBatchFile[i]);
970 952 : if (hashtable->outerBatchFile[i])
971 0 : BufFileClose(hashtable->outerBatchFile[i]);
972 : }
973 : }
974 :
975 : /* Release working memory (batchCxt is a child, so it goes away too) */
976 25378 : MemoryContextDelete(hashtable->hashCxt);
977 :
978 : /* And drop the control block */
979 25378 : pfree(hashtable);
980 25378 : }
981 :
982 : /*
983 : * Consider adjusting the allowed hash table size, depending on the number
984 : * of batches, to minimize the overall memory usage (for both the hashtable
985 : * and batch files).
986 : *
987 : * We're adjusting the size of the hash table, not the (optimal) number of
988 : * buckets. We can't change that once we start building the hash, due to how
989 : * ExecHashGetBucketAndBatch calculates batchno/bucketno from the hash. This
990 : * means the load factor may not be optimal, but we're in damage control so
991 : * we accept slower lookups. It's still much better than batch explosion.
992 : *
993 : * Returns true if we chose to increase the batch size (and thus we don't
994 : * need to add batches), and false if we should increase nbatch.
995 : */
996 : static bool
997 196 : ExecHashIncreaseBatchSize(HashJoinTable hashtable)
998 : {
999 : /*
1000 : * How much additional memory would doubling nbatch use? Each batch may
1001 : * require two buffered files (inner/outer), with a BLCKSZ buffer.
1002 : */
1003 196 : size_t batchSpace = (hashtable->nbatch * 2 * (size_t) BLCKSZ);
1004 :
1005 : /*
1006 : * Compare the new space needed for doubling nbatch and for enlarging the
1007 : * in-memory hash table. If doubling the hash table needs less memory,
1008 : * just do that. Otherwise, continue with doubling the nbatch.
1009 : *
1010 : * We're either doubling spaceAllowed or batchSpace, so which of those
1011 : * increases the memory usage the least is the same as comparing the
1012 : * values directly.
1013 : */
1014 196 : if (hashtable->spaceAllowed <= batchSpace)
1015 : {
1016 0 : hashtable->spaceAllowed *= 2;
1017 0 : return true;
1018 : }
1019 :
1020 196 : return false;
1021 : }
1022 :
1023 : /*
1024 : * ExecHashIncreaseNumBatches
1025 : * increase the original number of batches in order to reduce
1026 : * current memory consumption
1027 : */
1028 : static void
1029 794616 : ExecHashIncreaseNumBatches(HashJoinTable hashtable)
1030 : {
1031 794616 : int oldnbatch = hashtable->nbatch;
1032 794616 : int curbatch = hashtable->curbatch;
1033 : int nbatch;
1034 : long ninmemory;
1035 : long nfreed;
1036 : HashMemoryChunk oldchunks;
1037 :
1038 : /* do nothing if we've decided to shut off growth */
1039 794616 : if (!hashtable->growEnabled)
1040 794420 : return;
1041 :
1042 : /* safety check to avoid overflow */
1043 196 : if (oldnbatch > Min(INT_MAX / 2, MaxAllocSize / (sizeof(void *) * 2)))
1044 0 : return;
1045 :
1046 : /* consider increasing size of the in-memory hash table instead */
1047 196 : if (ExecHashIncreaseBatchSize(hashtable))
1048 0 : return;
1049 :
1050 196 : nbatch = oldnbatch * 2;
1051 : Assert(nbatch > 1);
1052 :
1053 : #ifdef HJDEBUG
1054 : printf("Hashjoin %p: increasing nbatch to %d because space = %zu\n",
1055 : hashtable, nbatch, hashtable->spaceUsed);
1056 : #endif
1057 :
1058 196 : if (hashtable->innerBatchFile == NULL)
1059 : {
1060 94 : MemoryContext oldcxt = MemoryContextSwitchTo(hashtable->spillCxt);
1061 :
1062 : /* we had no file arrays before */
1063 94 : hashtable->innerBatchFile = palloc0_array(BufFile *, nbatch);
1064 94 : hashtable->outerBatchFile = palloc0_array(BufFile *, nbatch);
1065 :
1066 94 : MemoryContextSwitchTo(oldcxt);
1067 :
1068 : /* time to establish the temp tablespaces, too */
1069 94 : PrepareTempTablespaces();
1070 : }
1071 : else
1072 : {
1073 : /* enlarge arrays and zero out added entries */
1074 102 : hashtable->innerBatchFile = repalloc0_array(hashtable->innerBatchFile, BufFile *, oldnbatch, nbatch);
1075 102 : hashtable->outerBatchFile = repalloc0_array(hashtable->outerBatchFile, BufFile *, oldnbatch, nbatch);
1076 : }
1077 :
1078 196 : hashtable->nbatch = nbatch;
1079 :
1080 : /*
1081 : * Scan through the existing hash table entries and dump out any that are
1082 : * no longer of the current batch.
1083 : */
1084 196 : ninmemory = nfreed = 0;
1085 :
1086 : /* If know we need to resize nbuckets, we can do it while rebatching. */
1087 196 : if (hashtable->nbuckets_optimal != hashtable->nbuckets)
1088 : {
1089 : /* we never decrease the number of buckets */
1090 : Assert(hashtable->nbuckets_optimal > hashtable->nbuckets);
1091 :
1092 94 : hashtable->nbuckets = hashtable->nbuckets_optimal;
1093 94 : hashtable->log2_nbuckets = hashtable->log2_nbuckets_optimal;
1094 :
1095 94 : hashtable->buckets.unshared =
1096 94 : repalloc_array(hashtable->buckets.unshared,
1097 : HashJoinTuple, hashtable->nbuckets);
1098 : }
1099 :
1100 : /*
1101 : * We will scan through the chunks directly, so that we can reset the
1102 : * buckets now and not have to keep track which tuples in the buckets have
1103 : * already been processed. We will free the old chunks as we go.
1104 : */
1105 196 : memset(hashtable->buckets.unshared, 0,
1106 196 : sizeof(HashJoinTuple) * hashtable->nbuckets);
1107 196 : oldchunks = hashtable->chunks;
1108 196 : hashtable->chunks = NULL;
1109 :
1110 : /* so, let's scan through the old chunks, and all tuples in each chunk */
1111 980 : while (oldchunks != NULL)
1112 : {
1113 784 : HashMemoryChunk nextchunk = oldchunks->next.unshared;
1114 :
1115 : /* position within the buffer (up to oldchunks->used) */
1116 784 : size_t idx = 0;
1117 :
1118 : /* process all tuples stored in this chunk (and then free it) */
1119 535628 : while (idx < oldchunks->used)
1120 : {
1121 534844 : HashJoinTuple hashTuple = (HashJoinTuple) (HASH_CHUNK_DATA(oldchunks) + idx);
1122 534844 : MinimalTuple tuple = HJTUPLE_MINTUPLE(hashTuple);
1123 534844 : int hashTupleSize = (HJTUPLE_OVERHEAD + tuple->t_len);
1124 : int bucketno;
1125 : int batchno;
1126 :
1127 534844 : ninmemory++;
1128 534844 : ExecHashGetBucketAndBatch(hashtable, hashTuple->hashvalue,
1129 : &bucketno, &batchno);
1130 :
1131 534844 : if (batchno == curbatch)
1132 : {
1133 : /* keep tuple in memory - copy it into the new chunk */
1134 : HashJoinTuple copyTuple;
1135 :
1136 202866 : copyTuple = (HashJoinTuple) dense_alloc(hashtable, hashTupleSize);
1137 202866 : memcpy(copyTuple, hashTuple, hashTupleSize);
1138 :
1139 : /* and add it back to the appropriate bucket */
1140 202866 : copyTuple->next.unshared = hashtable->buckets.unshared[bucketno];
1141 202866 : hashtable->buckets.unshared[bucketno] = copyTuple;
1142 : }
1143 : else
1144 : {
1145 : /* dump it out */
1146 : Assert(batchno > curbatch);
1147 331978 : ExecHashJoinSaveTuple(HJTUPLE_MINTUPLE(hashTuple),
1148 : hashTuple->hashvalue,
1149 331978 : &hashtable->innerBatchFile[batchno],
1150 : hashtable);
1151 :
1152 331978 : hashtable->spaceUsed -= hashTupleSize;
1153 331978 : nfreed++;
1154 : }
1155 :
1156 : /* next tuple in this chunk */
1157 534844 : idx += MAXALIGN(hashTupleSize);
1158 :
1159 : /* allow this loop to be cancellable */
1160 534844 : CHECK_FOR_INTERRUPTS();
1161 : }
1162 :
1163 : /* we're done with this chunk - free it and proceed to the next one */
1164 784 : pfree(oldchunks);
1165 784 : oldchunks = nextchunk;
1166 : }
1167 :
1168 : #ifdef HJDEBUG
1169 : printf("Hashjoin %p: freed %ld of %ld tuples, space now %zu\n",
1170 : hashtable, nfreed, ninmemory, hashtable->spaceUsed);
1171 : #endif
1172 :
1173 : /*
1174 : * If we dumped out either all or none of the tuples in the table, disable
1175 : * further expansion of nbatch. This situation implies that we have
1176 : * enough tuples of identical hashvalues to overflow spaceAllowed.
1177 : * Increasing nbatch will not fix it since there's no way to subdivide the
1178 : * group any more finely. We have to just gut it out and hope the server
1179 : * has enough RAM.
1180 : */
1181 196 : if (nfreed == 0 || nfreed == ninmemory)
1182 : {
1183 46 : hashtable->growEnabled = false;
1184 : #ifdef HJDEBUG
1185 : printf("Hashjoin %p: disabling further increase of nbatch\n",
1186 : hashtable);
1187 : #endif
1188 : }
1189 : }
1190 :
1191 : /*
1192 : * ExecParallelHashIncreaseNumBatches
1193 : * Every participant attached to grow_batches_barrier must run this
1194 : * function when it observes growth == PHJ_GROWTH_NEED_MORE_BATCHES.
1195 : */
1196 : static void
1197 54 : ExecParallelHashIncreaseNumBatches(HashJoinTable hashtable)
1198 : {
1199 54 : ParallelHashJoinState *pstate = hashtable->parallel_state;
1200 :
1201 : Assert(BarrierPhase(&pstate->build_barrier) == PHJ_BUILD_HASH_INNER);
1202 :
1203 : /*
1204 : * It's unlikely, but we need to be prepared for new participants to show
1205 : * up while we're in the middle of this operation so we need to switch on
1206 : * barrier phase here.
1207 : */
1208 54 : switch (PHJ_GROW_BATCHES_PHASE(BarrierPhase(&pstate->grow_batches_barrier)))
1209 : {
1210 54 : case PHJ_GROW_BATCHES_ELECT:
1211 :
1212 : /*
1213 : * Elect one participant to prepare to grow the number of batches.
1214 : * This involves reallocating or resetting the buckets of batch 0
1215 : * in preparation for all participants to begin repartitioning the
1216 : * tuples.
1217 : */
1218 54 : if (BarrierArriveAndWait(&pstate->grow_batches_barrier,
1219 : WAIT_EVENT_HASH_GROW_BATCHES_ELECT))
1220 : {
1221 : dsa_pointer_atomic *buckets;
1222 : ParallelHashJoinBatch *old_batch0;
1223 : int new_nbatch;
1224 : int i;
1225 :
1226 : /* Move the old batch out of the way. */
1227 52 : old_batch0 = hashtable->batches[0].shared;
1228 52 : pstate->old_batches = pstate->batches;
1229 52 : pstate->old_nbatch = hashtable->nbatch;
1230 52 : pstate->batches = InvalidDsaPointer;
1231 :
1232 : /* Free this backend's old accessors. */
1233 52 : ExecParallelHashCloseBatchAccessors(hashtable);
1234 :
1235 : /* Figure out how many batches to use. */
1236 52 : if (hashtable->nbatch == 1)
1237 : {
1238 : /*
1239 : * We are going from single-batch to multi-batch. We need
1240 : * to switch from one large combined memory budget to the
1241 : * regular hash_mem budget.
1242 : */
1243 36 : pstate->space_allowed = get_hash_memory_limit();
1244 :
1245 : /*
1246 : * The combined hash_mem of all participants wasn't
1247 : * enough. Therefore one batch per participant would be
1248 : * approximately equivalent and would probably also be
1249 : * insufficient. So try two batches per participant,
1250 : * rounded up to a power of two.
1251 : */
1252 36 : new_nbatch = pg_nextpower2_32(pstate->nparticipants * 2);
1253 : }
1254 : else
1255 : {
1256 : /*
1257 : * We were already multi-batched. Try doubling the number
1258 : * of batches.
1259 : */
1260 16 : new_nbatch = hashtable->nbatch * 2;
1261 : }
1262 :
1263 : /* Allocate new larger generation of batches. */
1264 : Assert(hashtable->nbatch == pstate->nbatch);
1265 52 : ExecParallelHashJoinSetUpBatches(hashtable, new_nbatch);
1266 : Assert(hashtable->nbatch == pstate->nbatch);
1267 :
1268 : /* Replace or recycle batch 0's bucket array. */
1269 52 : if (pstate->old_nbatch == 1)
1270 : {
1271 : double dtuples;
1272 : double dbuckets;
1273 : int new_nbuckets;
1274 : uint32 max_buckets;
1275 :
1276 : /*
1277 : * We probably also need a smaller bucket array. How many
1278 : * tuples do we expect per batch, assuming we have only
1279 : * half of them so far? Normally we don't need to change
1280 : * the bucket array's size, because the size of each batch
1281 : * stays the same as we add more batches, but in this
1282 : * special case we move from a large batch to many smaller
1283 : * batches and it would be wasteful to keep the large
1284 : * array.
1285 : */
1286 36 : dtuples = (old_batch0->ntuples * 2.0) / new_nbatch;
1287 :
1288 : /*
1289 : * We need to calculate the maximum number of buckets to
1290 : * stay within the MaxAllocSize boundary. Round the
1291 : * maximum number to the previous power of 2 given that
1292 : * later we round the number to the next power of 2.
1293 : */
1294 36 : max_buckets = pg_prevpower2_32((uint32)
1295 : (MaxAllocSize / sizeof(dsa_pointer_atomic)));
1296 36 : dbuckets = ceil(dtuples / NTUP_PER_BUCKET);
1297 36 : dbuckets = Min(dbuckets, max_buckets);
1298 36 : new_nbuckets = (int) dbuckets;
1299 36 : new_nbuckets = Max(new_nbuckets, 1024);
1300 36 : new_nbuckets = pg_nextpower2_32(new_nbuckets);
1301 36 : dsa_free(hashtable->area, old_batch0->buckets);
1302 72 : hashtable->batches[0].shared->buckets =
1303 36 : dsa_allocate(hashtable->area,
1304 : sizeof(dsa_pointer_atomic) * new_nbuckets);
1305 : buckets = (dsa_pointer_atomic *)
1306 36 : dsa_get_address(hashtable->area,
1307 36 : hashtable->batches[0].shared->buckets);
1308 110628 : for (i = 0; i < new_nbuckets; ++i)
1309 110592 : dsa_pointer_atomic_init(&buckets[i], InvalidDsaPointer);
1310 36 : pstate->nbuckets = new_nbuckets;
1311 : }
1312 : else
1313 : {
1314 : /* Recycle the existing bucket array. */
1315 16 : hashtable->batches[0].shared->buckets = old_batch0->buckets;
1316 : buckets = (dsa_pointer_atomic *)
1317 16 : dsa_get_address(hashtable->area, old_batch0->buckets);
1318 65552 : for (i = 0; i < hashtable->nbuckets; ++i)
1319 65536 : dsa_pointer_atomic_write(&buckets[i], InvalidDsaPointer);
1320 : }
1321 :
1322 : /* Move all chunks to the work queue for parallel processing. */
1323 52 : pstate->chunk_work_queue = old_batch0->chunks;
1324 :
1325 : /* Disable further growth temporarily while we're growing. */
1326 52 : pstate->growth = PHJ_GROWTH_DISABLED;
1327 : }
1328 : else
1329 : {
1330 : /* All other participants just flush their tuples to disk. */
1331 2 : ExecParallelHashCloseBatchAccessors(hashtable);
1332 : }
1333 : /* Fall through. */
1334 :
1335 : case PHJ_GROW_BATCHES_REALLOCATE:
1336 : /* Wait for the above to be finished. */
1337 54 : BarrierArriveAndWait(&pstate->grow_batches_barrier,
1338 : WAIT_EVENT_HASH_GROW_BATCHES_REALLOCATE);
1339 : /* Fall through. */
1340 :
1341 54 : case PHJ_GROW_BATCHES_REPARTITION:
1342 : /* Make sure that we have the current dimensions and buckets. */
1343 54 : ExecParallelHashEnsureBatchAccessors(hashtable);
1344 54 : ExecParallelHashTableSetCurrentBatch(hashtable, 0);
1345 : /* Then partition, flush counters. */
1346 54 : ExecParallelHashRepartitionFirst(hashtable);
1347 54 : ExecParallelHashRepartitionRest(hashtable);
1348 54 : ExecParallelHashMergeCounters(hashtable);
1349 : /* Wait for the above to be finished. */
1350 54 : BarrierArriveAndWait(&pstate->grow_batches_barrier,
1351 : WAIT_EVENT_HASH_GROW_BATCHES_REPARTITION);
1352 : /* Fall through. */
1353 :
1354 54 : case PHJ_GROW_BATCHES_DECIDE:
1355 :
1356 : /*
1357 : * Elect one participant to clean up and decide whether further
1358 : * repartitioning is needed, or should be disabled because it's
1359 : * not helping.
1360 : */
1361 54 : if (BarrierArriveAndWait(&pstate->grow_batches_barrier,
1362 : WAIT_EVENT_HASH_GROW_BATCHES_DECIDE))
1363 : {
1364 : ParallelHashJoinBatch *old_batches;
1365 52 : bool space_exhausted = false;
1366 52 : bool extreme_skew_detected = false;
1367 :
1368 : /* Make sure that we have the current dimensions and buckets. */
1369 52 : ExecParallelHashEnsureBatchAccessors(hashtable);
1370 52 : ExecParallelHashTableSetCurrentBatch(hashtable, 0);
1371 :
1372 52 : old_batches = dsa_get_address(hashtable->area, pstate->old_batches);
1373 :
1374 : /* Are any of the new generation of batches exhausted? */
1375 372 : for (int i = 0; i < hashtable->nbatch; ++i)
1376 : {
1377 : ParallelHashJoinBatch *batch;
1378 : ParallelHashJoinBatch *old_batch;
1379 : int parent;
1380 :
1381 320 : batch = hashtable->batches[i].shared;
1382 320 : if (batch->space_exhausted ||
1383 320 : batch->estimated_size > pstate->space_allowed)
1384 24 : space_exhausted = true;
1385 :
1386 320 : parent = i % pstate->old_nbatch;
1387 320 : old_batch = NthParallelHashJoinBatch(old_batches, parent);
1388 320 : if (old_batch->space_exhausted ||
1389 96 : batch->estimated_size > pstate->space_allowed)
1390 : {
1391 : /*
1392 : * Did this batch receive ALL of the tuples from its
1393 : * parent batch? That would indicate that further
1394 : * repartitioning isn't going to help (the hash values
1395 : * are probably all the same).
1396 : */
1397 224 : if (batch->ntuples == hashtable->batches[parent].shared->old_ntuples)
1398 24 : extreme_skew_detected = true;
1399 : }
1400 : }
1401 :
1402 : /* Don't keep growing if it's not helping or we'd overflow. */
1403 52 : if (extreme_skew_detected || hashtable->nbatch >= INT_MAX / 2)
1404 24 : pstate->growth = PHJ_GROWTH_DISABLED;
1405 28 : else if (space_exhausted)
1406 0 : pstate->growth = PHJ_GROWTH_NEED_MORE_BATCHES;
1407 : else
1408 28 : pstate->growth = PHJ_GROWTH_OK;
1409 :
1410 : /* Free the old batches in shared memory. */
1411 52 : dsa_free(hashtable->area, pstate->old_batches);
1412 52 : pstate->old_batches = InvalidDsaPointer;
1413 : }
1414 : /* Fall through. */
1415 :
1416 : case PHJ_GROW_BATCHES_FINISH:
1417 : /* Wait for the above to complete. */
1418 54 : BarrierArriveAndWait(&pstate->grow_batches_barrier,
1419 : WAIT_EVENT_HASH_GROW_BATCHES_FINISH);
1420 : }
1421 54 : }
1422 :
1423 : /*
1424 : * Repartition the tuples currently loaded into memory for inner batch 0
1425 : * because the number of batches has been increased. Some tuples are retained
1426 : * in memory and some are written out to a later batch.
1427 : */
1428 : static void
1429 54 : ExecParallelHashRepartitionFirst(HashJoinTable hashtable)
1430 : {
1431 : dsa_pointer chunk_shared;
1432 : HashMemoryChunk chunk;
1433 :
1434 : Assert(hashtable->nbatch == hashtable->parallel_state->nbatch);
1435 :
1436 410 : while ((chunk = ExecParallelHashPopChunkQueue(hashtable, &chunk_shared)))
1437 : {
1438 302 : size_t idx = 0;
1439 :
1440 : /* Repartition all tuples in this chunk. */
1441 230660 : while (idx < chunk->used)
1442 : {
1443 230358 : HashJoinTuple hashTuple = (HashJoinTuple) (HASH_CHUNK_DATA(chunk) + idx);
1444 230358 : MinimalTuple tuple = HJTUPLE_MINTUPLE(hashTuple);
1445 : HashJoinTuple copyTuple;
1446 : dsa_pointer shared;
1447 : int bucketno;
1448 : int batchno;
1449 :
1450 230358 : ExecHashGetBucketAndBatch(hashtable, hashTuple->hashvalue,
1451 : &bucketno, &batchno);
1452 :
1453 : Assert(batchno < hashtable->nbatch);
1454 230358 : if (batchno == 0)
1455 : {
1456 : /* It still belongs in batch 0. Copy to a new chunk. */
1457 : copyTuple =
1458 55578 : ExecParallelHashTupleAlloc(hashtable,
1459 55578 : HJTUPLE_OVERHEAD + tuple->t_len,
1460 : &shared);
1461 55578 : copyTuple->hashvalue = hashTuple->hashvalue;
1462 55578 : memcpy(HJTUPLE_MINTUPLE(copyTuple), tuple, tuple->t_len);
1463 55578 : ExecParallelHashPushTuple(&hashtable->buckets.shared[bucketno],
1464 : copyTuple, shared);
1465 : }
1466 : else
1467 : {
1468 174780 : size_t tuple_size =
1469 174780 : MAXALIGN(HJTUPLE_OVERHEAD + tuple->t_len);
1470 :
1471 : /* It belongs in a later batch. */
1472 174780 : hashtable->batches[batchno].estimated_size += tuple_size;
1473 174780 : sts_puttuple(hashtable->batches[batchno].inner_tuples,
1474 174780 : &hashTuple->hashvalue, tuple);
1475 : }
1476 :
1477 : /* Count this tuple. */
1478 230358 : ++hashtable->batches[0].old_ntuples;
1479 230358 : ++hashtable->batches[batchno].ntuples;
1480 :
1481 230358 : idx += MAXALIGN(HJTUPLE_OVERHEAD +
1482 : HJTUPLE_MINTUPLE(hashTuple)->t_len);
1483 : }
1484 :
1485 : /* Free this chunk. */
1486 302 : dsa_free(hashtable->area, chunk_shared);
1487 :
1488 302 : CHECK_FOR_INTERRUPTS();
1489 : }
1490 54 : }
1491 :
1492 : /*
1493 : * Help repartition inner batches 1..n.
1494 : */
1495 : static void
1496 54 : ExecParallelHashRepartitionRest(HashJoinTable hashtable)
1497 : {
1498 54 : ParallelHashJoinState *pstate = hashtable->parallel_state;
1499 54 : int old_nbatch = pstate->old_nbatch;
1500 : SharedTuplestoreAccessor **old_inner_tuples;
1501 : ParallelHashJoinBatch *old_batches;
1502 : int i;
1503 :
1504 : /* Get our hands on the previous generation of batches. */
1505 : old_batches = (ParallelHashJoinBatch *)
1506 54 : dsa_get_address(hashtable->area, pstate->old_batches);
1507 54 : old_inner_tuples = palloc0_array(SharedTuplestoreAccessor *, old_nbatch);
1508 108 : for (i = 1; i < old_nbatch; ++i)
1509 : {
1510 54 : ParallelHashJoinBatch *shared =
1511 54 : NthParallelHashJoinBatch(old_batches, i);
1512 :
1513 54 : old_inner_tuples[i] = sts_attach(ParallelHashJoinBatchInner(shared),
1514 : ParallelWorkerNumber + 1,
1515 : &pstate->fileset);
1516 : }
1517 :
1518 : /* Join in the effort to repartition them. */
1519 108 : for (i = 1; i < old_nbatch; ++i)
1520 : {
1521 : MinimalTuple tuple;
1522 : uint32 hashvalue;
1523 :
1524 : /* Scan one partition from the previous generation. */
1525 54 : sts_begin_parallel_scan(old_inner_tuples[i]);
1526 191040 : while ((tuple = sts_parallel_scan_next(old_inner_tuples[i], &hashvalue)))
1527 : {
1528 190986 : size_t tuple_size = MAXALIGN(HJTUPLE_OVERHEAD + tuple->t_len);
1529 : int bucketno;
1530 : int batchno;
1531 :
1532 : /* Decide which partition it goes to in the new generation. */
1533 190986 : ExecHashGetBucketAndBatch(hashtable, hashvalue, &bucketno,
1534 : &batchno);
1535 :
1536 190986 : hashtable->batches[batchno].estimated_size += tuple_size;
1537 190986 : ++hashtable->batches[batchno].ntuples;
1538 190986 : ++hashtable->batches[i].old_ntuples;
1539 :
1540 : /* Store the tuple its new batch. */
1541 190986 : sts_puttuple(hashtable->batches[batchno].inner_tuples,
1542 : &hashvalue, tuple);
1543 :
1544 190986 : CHECK_FOR_INTERRUPTS();
1545 : }
1546 54 : sts_end_parallel_scan(old_inner_tuples[i]);
1547 : }
1548 :
1549 54 : pfree(old_inner_tuples);
1550 54 : }
1551 :
1552 : /*
1553 : * Transfer the backend-local per-batch counters to the shared totals.
1554 : */
1555 : static void
1556 344 : ExecParallelHashMergeCounters(HashJoinTable hashtable)
1557 : {
1558 344 : ParallelHashJoinState *pstate = hashtable->parallel_state;
1559 : int i;
1560 :
1561 344 : LWLockAcquire(&pstate->lock, LW_EXCLUSIVE);
1562 344 : pstate->total_tuples = 0;
1563 1686 : for (i = 0; i < hashtable->nbatch; ++i)
1564 : {
1565 1342 : ParallelHashJoinBatchAccessor *batch = &hashtable->batches[i];
1566 :
1567 1342 : batch->shared->size += batch->size;
1568 1342 : batch->shared->estimated_size += batch->estimated_size;
1569 1342 : batch->shared->ntuples += batch->ntuples;
1570 1342 : batch->shared->old_ntuples += batch->old_ntuples;
1571 1342 : batch->size = 0;
1572 1342 : batch->estimated_size = 0;
1573 1342 : batch->ntuples = 0;
1574 1342 : batch->old_ntuples = 0;
1575 1342 : pstate->total_tuples += batch->shared->ntuples;
1576 : }
1577 344 : LWLockRelease(&pstate->lock);
1578 344 : }
1579 :
1580 : /*
1581 : * ExecHashIncreaseNumBuckets
1582 : * increase the original number of buckets in order to reduce
1583 : * number of tuples per bucket
1584 : */
1585 : static void
1586 72 : ExecHashIncreaseNumBuckets(HashJoinTable hashtable)
1587 : {
1588 : HashMemoryChunk chunk;
1589 :
1590 : /* do nothing if not an increase (it's called increase for a reason) */
1591 72 : if (hashtable->nbuckets >= hashtable->nbuckets_optimal)
1592 0 : return;
1593 :
1594 : #ifdef HJDEBUG
1595 : printf("Hashjoin %p: increasing nbuckets %d => %d\n",
1596 : hashtable, hashtable->nbuckets, hashtable->nbuckets_optimal);
1597 : #endif
1598 :
1599 72 : hashtable->nbuckets = hashtable->nbuckets_optimal;
1600 72 : hashtable->log2_nbuckets = hashtable->log2_nbuckets_optimal;
1601 :
1602 : Assert(hashtable->nbuckets > 1);
1603 : Assert(hashtable->nbuckets <= (INT_MAX / 2));
1604 : Assert(hashtable->nbuckets == (1 << hashtable->log2_nbuckets));
1605 :
1606 : /*
1607 : * Just reallocate the proper number of buckets - we don't need to walk
1608 : * through them - we can walk the dense-allocated chunks (just like in
1609 : * ExecHashIncreaseNumBatches, but without all the copying into new
1610 : * chunks)
1611 : */
1612 72 : hashtable->buckets.unshared =
1613 72 : repalloc_array(hashtable->buckets.unshared,
1614 : HashJoinTuple, hashtable->nbuckets);
1615 :
1616 72 : memset(hashtable->buckets.unshared, 0,
1617 72 : hashtable->nbuckets * sizeof(HashJoinTuple));
1618 :
1619 : /* scan through all tuples in all chunks to rebuild the hash table */
1620 1008 : for (chunk = hashtable->chunks; chunk != NULL; chunk = chunk->next.unshared)
1621 : {
1622 : /* process all tuples stored in this chunk */
1623 936 : size_t idx = 0;
1624 :
1625 720936 : while (idx < chunk->used)
1626 : {
1627 720000 : HashJoinTuple hashTuple = (HashJoinTuple) (HASH_CHUNK_DATA(chunk) + idx);
1628 : int bucketno;
1629 : int batchno;
1630 :
1631 720000 : ExecHashGetBucketAndBatch(hashtable, hashTuple->hashvalue,
1632 : &bucketno, &batchno);
1633 :
1634 : /* add the tuple to the proper bucket */
1635 720000 : hashTuple->next.unshared = hashtable->buckets.unshared[bucketno];
1636 720000 : hashtable->buckets.unshared[bucketno] = hashTuple;
1637 :
1638 : /* advance index past the tuple */
1639 720000 : idx += MAXALIGN(HJTUPLE_OVERHEAD +
1640 : HJTUPLE_MINTUPLE(hashTuple)->t_len);
1641 : }
1642 :
1643 : /* allow this loop to be cancellable */
1644 936 : CHECK_FOR_INTERRUPTS();
1645 : }
1646 : }
1647 :
1648 : static void
1649 126 : ExecParallelHashIncreaseNumBuckets(HashJoinTable hashtable)
1650 : {
1651 126 : ParallelHashJoinState *pstate = hashtable->parallel_state;
1652 : int i;
1653 : HashMemoryChunk chunk;
1654 : dsa_pointer chunk_s;
1655 :
1656 : Assert(BarrierPhase(&pstate->build_barrier) == PHJ_BUILD_HASH_INNER);
1657 :
1658 : /*
1659 : * It's unlikely, but we need to be prepared for new participants to show
1660 : * up while we're in the middle of this operation so we need to switch on
1661 : * barrier phase here.
1662 : */
1663 126 : switch (PHJ_GROW_BUCKETS_PHASE(BarrierPhase(&pstate->grow_buckets_barrier)))
1664 : {
1665 126 : case PHJ_GROW_BUCKETS_ELECT:
1666 : /* Elect one participant to prepare to increase nbuckets. */
1667 126 : if (BarrierArriveAndWait(&pstate->grow_buckets_barrier,
1668 : WAIT_EVENT_HASH_GROW_BUCKETS_ELECT))
1669 : {
1670 : size_t size;
1671 : dsa_pointer_atomic *buckets;
1672 :
1673 : /* Double the size of the bucket array. */
1674 108 : pstate->nbuckets *= 2;
1675 108 : size = pstate->nbuckets * sizeof(dsa_pointer_atomic);
1676 108 : hashtable->batches[0].shared->size += size / 2;
1677 108 : dsa_free(hashtable->area, hashtable->batches[0].shared->buckets);
1678 216 : hashtable->batches[0].shared->buckets =
1679 108 : dsa_allocate(hashtable->area, size);
1680 : buckets = (dsa_pointer_atomic *)
1681 108 : dsa_get_address(hashtable->area,
1682 108 : hashtable->batches[0].shared->buckets);
1683 933996 : for (i = 0; i < pstate->nbuckets; ++i)
1684 933888 : dsa_pointer_atomic_init(&buckets[i], InvalidDsaPointer);
1685 :
1686 : /* Put the chunk list onto the work queue. */
1687 108 : pstate->chunk_work_queue = hashtable->batches[0].shared->chunks;
1688 :
1689 : /* Clear the flag. */
1690 108 : pstate->growth = PHJ_GROWTH_OK;
1691 : }
1692 : /* Fall through. */
1693 :
1694 : case PHJ_GROW_BUCKETS_REALLOCATE:
1695 : /* Wait for the above to complete. */
1696 126 : BarrierArriveAndWait(&pstate->grow_buckets_barrier,
1697 : WAIT_EVENT_HASH_GROW_BUCKETS_REALLOCATE);
1698 : /* Fall through. */
1699 :
1700 126 : case PHJ_GROW_BUCKETS_REINSERT:
1701 : /* Reinsert all tuples into the hash table. */
1702 126 : ExecParallelHashEnsureBatchAccessors(hashtable);
1703 126 : ExecParallelHashTableSetCurrentBatch(hashtable, 0);
1704 934 : while ((chunk = ExecParallelHashPopChunkQueue(hashtable, &chunk_s)))
1705 : {
1706 682 : size_t idx = 0;
1707 :
1708 557068 : while (idx < chunk->used)
1709 : {
1710 556386 : HashJoinTuple hashTuple = (HashJoinTuple) (HASH_CHUNK_DATA(chunk) + idx);
1711 556386 : dsa_pointer shared = chunk_s + HASH_CHUNK_HEADER_SIZE + idx;
1712 : int bucketno;
1713 : int batchno;
1714 :
1715 556386 : ExecHashGetBucketAndBatch(hashtable, hashTuple->hashvalue,
1716 : &bucketno, &batchno);
1717 : Assert(batchno == 0);
1718 :
1719 : /* add the tuple to the proper bucket */
1720 556386 : ExecParallelHashPushTuple(&hashtable->buckets.shared[bucketno],
1721 : hashTuple, shared);
1722 :
1723 : /* advance index past the tuple */
1724 556386 : idx += MAXALIGN(HJTUPLE_OVERHEAD +
1725 : HJTUPLE_MINTUPLE(hashTuple)->t_len);
1726 : }
1727 :
1728 : /* allow this loop to be cancellable */
1729 682 : CHECK_FOR_INTERRUPTS();
1730 : }
1731 126 : BarrierArriveAndWait(&pstate->grow_buckets_barrier,
1732 : WAIT_EVENT_HASH_GROW_BUCKETS_REINSERT);
1733 : }
1734 126 : }
1735 :
1736 : /*
1737 : * ExecHashTableInsert
1738 : * insert a tuple into the hash table depending on the hash value
1739 : * it may just go to a temp file for later batches
1740 : *
1741 : * Note: the passed TupleTableSlot may contain a regular, minimal, or virtual
1742 : * tuple; the minimal case in particular is certain to happen while reloading
1743 : * tuples from batch files. We could save some cycles in the regular-tuple
1744 : * case by not forcing the slot contents into minimal form; not clear if it's
1745 : * worth the messiness required.
1746 : */
1747 : void
1748 12328468 : ExecHashTableInsert(HashJoinTable hashtable,
1749 : TupleTableSlot *slot,
1750 : uint32 hashvalue)
1751 : {
1752 : bool shouldFree;
1753 12328468 : MinimalTuple tuple = ExecFetchSlotMinimalTuple(slot, &shouldFree);
1754 : int bucketno;
1755 : int batchno;
1756 :
1757 12328468 : ExecHashGetBucketAndBatch(hashtable, hashvalue,
1758 : &bucketno, &batchno);
1759 :
1760 : /*
1761 : * decide whether to put the tuple in the hash table or a temp file
1762 : */
1763 12328468 : if (batchno == hashtable->curbatch)
1764 : {
1765 : /*
1766 : * put the tuple in hash table
1767 : */
1768 : HashJoinTuple hashTuple;
1769 : int hashTupleSize;
1770 9235986 : double ntuples = (hashtable->totalTuples - hashtable->skewTuples);
1771 :
1772 : /* Create the HashJoinTuple */
1773 9235986 : hashTupleSize = HJTUPLE_OVERHEAD + tuple->t_len;
1774 9235986 : hashTuple = (HashJoinTuple) dense_alloc(hashtable, hashTupleSize);
1775 :
1776 9235986 : hashTuple->hashvalue = hashvalue;
1777 9235986 : memcpy(HJTUPLE_MINTUPLE(hashTuple), tuple, tuple->t_len);
1778 :
1779 : /*
1780 : * We always reset the tuple-matched flag on insertion. This is okay
1781 : * even when reloading a tuple from a batch file, since the tuple
1782 : * could not possibly have been matched to an outer tuple before it
1783 : * went into the batch file.
1784 : */
1785 9235986 : HeapTupleHeaderClearMatch(HJTUPLE_MINTUPLE(hashTuple));
1786 :
1787 : /* Push it onto the front of the bucket's list */
1788 9235986 : hashTuple->next.unshared = hashtable->buckets.unshared[bucketno];
1789 9235986 : hashtable->buckets.unshared[bucketno] = hashTuple;
1790 :
1791 : /*
1792 : * Increase the (optimal) number of buckets if we just exceeded the
1793 : * NTUP_PER_BUCKET threshold, but only when there's still a single
1794 : * batch.
1795 : */
1796 9235986 : if (hashtable->nbatch == 1 &&
1797 5541100 : ntuples > (hashtable->nbuckets_optimal * NTUP_PER_BUCKET))
1798 : {
1799 : /* Guard against integer overflow and alloc size overflow */
1800 260 : if (hashtable->nbuckets_optimal <= INT_MAX / 2 &&
1801 260 : hashtable->nbuckets_optimal * 2 <= MaxAllocSize / sizeof(HashJoinTuple))
1802 : {
1803 260 : hashtable->nbuckets_optimal *= 2;
1804 260 : hashtable->log2_nbuckets_optimal += 1;
1805 : }
1806 : }
1807 :
1808 : /* Account for space used, and back off if we've used too much */
1809 9235986 : hashtable->spaceUsed += hashTupleSize;
1810 9235986 : if (hashtable->spaceUsed > hashtable->spacePeak)
1811 6651704 : hashtable->spacePeak = hashtable->spaceUsed;
1812 9235986 : if (hashtable->spaceUsed +
1813 9235986 : hashtable->nbuckets_optimal * sizeof(HashJoinTuple)
1814 9235986 : > hashtable->spaceAllowed)
1815 794616 : ExecHashIncreaseNumBatches(hashtable);
1816 : }
1817 : else
1818 : {
1819 : /*
1820 : * put the tuple into a temp file for later batches
1821 : */
1822 : Assert(batchno > hashtable->curbatch);
1823 3092482 : ExecHashJoinSaveTuple(tuple,
1824 : hashvalue,
1825 3092482 : &hashtable->innerBatchFile[batchno],
1826 : hashtable);
1827 : }
1828 :
1829 12328468 : if (shouldFree)
1830 8827448 : heap_free_minimal_tuple(tuple);
1831 12328468 : }
1832 :
1833 : /*
1834 : * ExecParallelHashTableInsert
1835 : * insert a tuple into a shared hash table or shared batch tuplestore
1836 : */
1837 : void
1838 2160192 : ExecParallelHashTableInsert(HashJoinTable hashtable,
1839 : TupleTableSlot *slot,
1840 : uint32 hashvalue)
1841 : {
1842 : bool shouldFree;
1843 2160192 : MinimalTuple tuple = ExecFetchSlotMinimalTuple(slot, &shouldFree);
1844 : dsa_pointer shared;
1845 : int bucketno;
1846 : int batchno;
1847 :
1848 340 : retry:
1849 2160532 : ExecHashGetBucketAndBatch(hashtable, hashvalue, &bucketno, &batchno);
1850 :
1851 2160532 : if (batchno == 0)
1852 : {
1853 : HashJoinTuple hashTuple;
1854 :
1855 : /* Try to load it into memory. */
1856 : Assert(BarrierPhase(&hashtable->parallel_state->build_barrier) ==
1857 : PHJ_BUILD_HASH_INNER);
1858 1293834 : hashTuple = ExecParallelHashTupleAlloc(hashtable,
1859 1293834 : HJTUPLE_OVERHEAD + tuple->t_len,
1860 : &shared);
1861 1293834 : if (hashTuple == NULL)
1862 306 : goto retry;
1863 :
1864 : /* Store the hash value in the HashJoinTuple header. */
1865 1293528 : hashTuple->hashvalue = hashvalue;
1866 1293528 : memcpy(HJTUPLE_MINTUPLE(hashTuple), tuple, tuple->t_len);
1867 1293528 : HeapTupleHeaderClearMatch(HJTUPLE_MINTUPLE(hashTuple));
1868 :
1869 : /* Push it onto the front of the bucket's list */
1870 1293528 : ExecParallelHashPushTuple(&hashtable->buckets.shared[bucketno],
1871 : hashTuple, shared);
1872 : }
1873 : else
1874 : {
1875 866698 : size_t tuple_size = MAXALIGN(HJTUPLE_OVERHEAD + tuple->t_len);
1876 :
1877 : Assert(batchno > 0);
1878 :
1879 : /* Try to preallocate space in the batch if necessary. */
1880 866698 : if (hashtable->batches[batchno].preallocated < tuple_size)
1881 : {
1882 1516 : if (!ExecParallelHashTuplePrealloc(hashtable, batchno, tuple_size))
1883 34 : goto retry;
1884 : }
1885 :
1886 : Assert(hashtable->batches[batchno].preallocated >= tuple_size);
1887 866664 : hashtable->batches[batchno].preallocated -= tuple_size;
1888 866664 : sts_puttuple(hashtable->batches[batchno].inner_tuples, &hashvalue,
1889 : tuple);
1890 : }
1891 2160192 : ++hashtable->batches[batchno].ntuples;
1892 :
1893 2160192 : if (shouldFree)
1894 2160192 : heap_free_minimal_tuple(tuple);
1895 2160192 : }
1896 :
1897 : /*
1898 : * Insert a tuple into the current hash table. Unlike
1899 : * ExecParallelHashTableInsert, this version is not prepared to send the tuple
1900 : * to other batches or to run out of memory, and should only be called with
1901 : * tuples that belong in the current batch once growth has been disabled.
1902 : */
1903 : void
1904 1041444 : ExecParallelHashTableInsertCurrentBatch(HashJoinTable hashtable,
1905 : TupleTableSlot *slot,
1906 : uint32 hashvalue)
1907 : {
1908 : bool shouldFree;
1909 1041444 : MinimalTuple tuple = ExecFetchSlotMinimalTuple(slot, &shouldFree);
1910 : HashJoinTuple hashTuple;
1911 : dsa_pointer shared;
1912 : int batchno;
1913 : int bucketno;
1914 :
1915 1041444 : ExecHashGetBucketAndBatch(hashtable, hashvalue, &bucketno, &batchno);
1916 : Assert(batchno == hashtable->curbatch);
1917 1041444 : hashTuple = ExecParallelHashTupleAlloc(hashtable,
1918 1041444 : HJTUPLE_OVERHEAD + tuple->t_len,
1919 : &shared);
1920 1041444 : hashTuple->hashvalue = hashvalue;
1921 1041444 : memcpy(HJTUPLE_MINTUPLE(hashTuple), tuple, tuple->t_len);
1922 1041444 : HeapTupleHeaderClearMatch(HJTUPLE_MINTUPLE(hashTuple));
1923 1041444 : ExecParallelHashPushTuple(&hashtable->buckets.shared[bucketno],
1924 : hashTuple, shared);
1925 :
1926 1041444 : if (shouldFree)
1927 0 : heap_free_minimal_tuple(tuple);
1928 1041444 : }
1929 :
1930 :
1931 : /*
1932 : * ExecHashGetBucketAndBatch
1933 : * Determine the bucket number and batch number for a hash value
1934 : *
1935 : * Note: on-the-fly increases of nbatch must not change the bucket number
1936 : * for a given hash code (since we don't move tuples to different hash
1937 : * chains), and must only cause the batch number to remain the same or
1938 : * increase. Our algorithm is
1939 : * bucketno = hashvalue MOD nbuckets
1940 : * batchno = ROR(hashvalue, log2_nbuckets) MOD nbatch
1941 : * where nbuckets and nbatch are both expected to be powers of 2, so we can
1942 : * do the computations by shifting and masking. (This assumes that all hash
1943 : * functions are good about randomizing all their output bits, else we are
1944 : * likely to have very skewed bucket or batch occupancy.)
1945 : *
1946 : * nbuckets and log2_nbuckets may change while nbatch == 1 because of dynamic
1947 : * bucket count growth. Once we start batching, the value is fixed and does
1948 : * not change over the course of the join (making it possible to compute batch
1949 : * number the way we do here).
1950 : *
1951 : * nbatch is always a power of 2; we increase it only by doubling it. This
1952 : * effectively adds one more bit to the top of the batchno. In very large
1953 : * joins, we might run out of bits to add, so we do this by rotating the hash
1954 : * value. This causes batchno to steal bits from bucketno when the number of
1955 : * virtual buckets exceeds 2^32. It's better to have longer bucket chains
1956 : * than to lose the ability to divide batches.
1957 : */
1958 : void
1959 39995326 : ExecHashGetBucketAndBatch(HashJoinTable hashtable,
1960 : uint32 hashvalue,
1961 : int *bucketno,
1962 : int *batchno)
1963 : {
1964 39995326 : uint32 nbuckets = (uint32) hashtable->nbuckets;
1965 39995326 : uint32 nbatch = (uint32) hashtable->nbatch;
1966 :
1967 39995326 : if (nbatch > 1)
1968 : {
1969 15471356 : *bucketno = hashvalue & (nbuckets - 1);
1970 15471356 : *batchno = pg_rotate_right32(hashvalue,
1971 15471356 : hashtable->log2_nbuckets) & (nbatch - 1);
1972 : }
1973 : else
1974 : {
1975 24523970 : *bucketno = hashvalue & (nbuckets - 1);
1976 24523970 : *batchno = 0;
1977 : }
1978 39995326 : }
1979 :
1980 : /*
1981 : * ExecScanHashBucket
1982 : * scan a hash bucket for matches to the current outer tuple
1983 : *
1984 : * The current outer tuple must be stored in econtext->ecxt_outertuple.
1985 : *
1986 : * On success, the inner tuple is stored into hjstate->hj_CurTuple and
1987 : * econtext->ecxt_innertuple, using hjstate->hj_HashTupleSlot as the slot
1988 : * for the latter.
1989 : */
1990 : bool
1991 22578508 : ExecScanHashBucket(HashJoinState *hjstate,
1992 : ExprContext *econtext)
1993 : {
1994 22578508 : ExprState *hjclauses = hjstate->hashclauses;
1995 22578508 : HashJoinTable hashtable = hjstate->hj_HashTable;
1996 22578508 : HashJoinTuple hashTuple = hjstate->hj_CurTuple;
1997 22578508 : uint32 hashvalue = hjstate->hj_CurHashValue;
1998 :
1999 : /*
2000 : * hj_CurTuple is the address of the tuple last returned from the current
2001 : * bucket, or NULL if it's time to start scanning a new bucket.
2002 : *
2003 : * If the tuple hashed to a skew bucket then scan the skew bucket
2004 : * otherwise scan the standard hashtable bucket.
2005 : */
2006 22578508 : if (hashTuple != NULL)
2007 5105388 : hashTuple = hashTuple->next.unshared;
2008 17473120 : else if (hjstate->hj_CurSkewBucketNo != INVALID_SKEW_BUCKET_NO)
2009 2400 : hashTuple = hashtable->skewBucket[hjstate->hj_CurSkewBucketNo]->tuples;
2010 : else
2011 17470720 : hashTuple = hashtable->buckets.unshared[hjstate->hj_CurBucketNo];
2012 :
2013 26974138 : while (hashTuple != NULL)
2014 : {
2015 14844380 : if (hashTuple->hashvalue == hashvalue)
2016 : {
2017 : TupleTableSlot *inntuple;
2018 :
2019 : /* insert hashtable's tuple into exec slot so ExecQual sees it */
2020 10448762 : inntuple = ExecStoreMinimalTuple(HJTUPLE_MINTUPLE(hashTuple),
2021 : hjstate->hj_HashTupleSlot,
2022 : false); /* do not pfree */
2023 10448762 : econtext->ecxt_innertuple = inntuple;
2024 :
2025 10448762 : if (ExecQualAndReset(hjclauses, econtext))
2026 : {
2027 10448750 : hjstate->hj_CurTuple = hashTuple;
2028 10448750 : return true;
2029 : }
2030 : }
2031 :
2032 4395630 : hashTuple = hashTuple->next.unshared;
2033 : }
2034 :
2035 : /*
2036 : * no match
2037 : */
2038 12129758 : return false;
2039 : }
2040 :
2041 : /*
2042 : * ExecParallelScanHashBucket
2043 : * scan a hash bucket for matches to the current outer tuple
2044 : *
2045 : * The current outer tuple must be stored in econtext->ecxt_outertuple.
2046 : *
2047 : * On success, the inner tuple is stored into hjstate->hj_CurTuple and
2048 : * econtext->ecxt_innertuple, using hjstate->hj_HashTupleSlot as the slot
2049 : * for the latter.
2050 : */
2051 : bool
2052 4206108 : ExecParallelScanHashBucket(HashJoinState *hjstate,
2053 : ExprContext *econtext)
2054 : {
2055 4206108 : ExprState *hjclauses = hjstate->hashclauses;
2056 4206108 : HashJoinTable hashtable = hjstate->hj_HashTable;
2057 4206108 : HashJoinTuple hashTuple = hjstate->hj_CurTuple;
2058 4206108 : uint32 hashvalue = hjstate->hj_CurHashValue;
2059 :
2060 : /*
2061 : * hj_CurTuple is the address of the tuple last returned from the current
2062 : * bucket, or NULL if it's time to start scanning a new bucket.
2063 : */
2064 4206108 : if (hashTuple != NULL)
2065 2040078 : hashTuple = ExecParallelHashNextTuple(hashtable, hashTuple);
2066 : else
2067 2166030 : hashTuple = ExecParallelHashFirstTuple(hashtable,
2068 : hjstate->hj_CurBucketNo);
2069 :
2070 5600618 : while (hashTuple != NULL)
2071 : {
2072 3434588 : if (hashTuple->hashvalue == hashvalue)
2073 : {
2074 : TupleTableSlot *inntuple;
2075 :
2076 : /* insert hashtable's tuple into exec slot so ExecQual sees it */
2077 2040078 : inntuple = ExecStoreMinimalTuple(HJTUPLE_MINTUPLE(hashTuple),
2078 : hjstate->hj_HashTupleSlot,
2079 : false); /* do not pfree */
2080 2040078 : econtext->ecxt_innertuple = inntuple;
2081 :
2082 2040078 : if (ExecQualAndReset(hjclauses, econtext))
2083 : {
2084 2040078 : hjstate->hj_CurTuple = hashTuple;
2085 2040078 : return true;
2086 : }
2087 : }
2088 :
2089 1394510 : hashTuple = ExecParallelHashNextTuple(hashtable, hashTuple);
2090 : }
2091 :
2092 : /*
2093 : * no match
2094 : */
2095 2166030 : return false;
2096 : }
2097 :
2098 : /*
2099 : * ExecPrepHashTableForUnmatched
2100 : * set up for a series of ExecScanHashTableForUnmatched calls
2101 : */
2102 : void
2103 3996 : ExecPrepHashTableForUnmatched(HashJoinState *hjstate)
2104 : {
2105 : /*----------
2106 : * During this scan we use the HashJoinState fields as follows:
2107 : *
2108 : * hj_CurBucketNo: next regular bucket to scan
2109 : * hj_CurSkewBucketNo: next skew bucket (an index into skewBucketNums)
2110 : * hj_CurTuple: last tuple returned, or NULL to start next bucket
2111 : *----------
2112 : */
2113 3996 : hjstate->hj_CurBucketNo = 0;
2114 3996 : hjstate->hj_CurSkewBucketNo = 0;
2115 3996 : hjstate->hj_CurTuple = NULL;
2116 3996 : }
2117 :
2118 : /*
2119 : * Decide if this process is allowed to run the unmatched scan. If so, the
2120 : * batch barrier is advanced to PHJ_BATCH_SCAN and true is returned.
2121 : * Otherwise the batch is detached and false is returned.
2122 : */
2123 : bool
2124 96 : ExecParallelPrepHashTableForUnmatched(HashJoinState *hjstate)
2125 : {
2126 96 : HashJoinTable hashtable = hjstate->hj_HashTable;
2127 96 : int curbatch = hashtable->curbatch;
2128 96 : ParallelHashJoinBatch *batch = hashtable->batches[curbatch].shared;
2129 :
2130 : Assert(BarrierPhase(&batch->batch_barrier) == PHJ_BATCH_PROBE);
2131 :
2132 : /*
2133 : * It would not be deadlock-free to wait on the batch barrier, because it
2134 : * is in PHJ_BATCH_PROBE phase, and thus processes attached to it have
2135 : * already emitted tuples. Therefore, we'll hold a wait-free election:
2136 : * only one process can continue to the next phase, and all others detach
2137 : * from this batch. They can still go any work on other batches, if there
2138 : * are any.
2139 : */
2140 96 : if (!BarrierArriveAndDetachExceptLast(&batch->batch_barrier))
2141 : {
2142 : /* This process considers the batch to be done. */
2143 30 : hashtable->batches[hashtable->curbatch].done = true;
2144 :
2145 : /* Make sure any temporary files are closed. */
2146 30 : sts_end_parallel_scan(hashtable->batches[curbatch].inner_tuples);
2147 30 : sts_end_parallel_scan(hashtable->batches[curbatch].outer_tuples);
2148 :
2149 : /*
2150 : * Track largest batch we've seen, which would normally happen in
2151 : * ExecHashTableDetachBatch().
2152 : */
2153 30 : hashtable->spacePeak =
2154 30 : Max(hashtable->spacePeak,
2155 : batch->size + sizeof(dsa_pointer_atomic) * hashtable->nbuckets);
2156 30 : hashtable->curbatch = -1;
2157 30 : return false;
2158 : }
2159 :
2160 : /* Now we are alone with this batch. */
2161 : Assert(BarrierPhase(&batch->batch_barrier) == PHJ_BATCH_SCAN);
2162 :
2163 : /*
2164 : * Has another process decided to give up early and command all processes
2165 : * to skip the unmatched scan?
2166 : */
2167 66 : if (batch->skip_unmatched)
2168 : {
2169 0 : hashtable->batches[hashtable->curbatch].done = true;
2170 0 : ExecHashTableDetachBatch(hashtable);
2171 0 : return false;
2172 : }
2173 :
2174 : /* Now prepare the process local state, just as for non-parallel join. */
2175 66 : ExecPrepHashTableForUnmatched(hjstate);
2176 :
2177 66 : return true;
2178 : }
2179 :
2180 : /*
2181 : * ExecScanHashTableForUnmatched
2182 : * scan the hash table for unmatched inner tuples
2183 : *
2184 : * On success, the inner tuple is stored into hjstate->hj_CurTuple and
2185 : * econtext->ecxt_innertuple, using hjstate->hj_HashTupleSlot as the slot
2186 : * for the latter.
2187 : */
2188 : bool
2189 433524 : ExecScanHashTableForUnmatched(HashJoinState *hjstate, ExprContext *econtext)
2190 : {
2191 433524 : HashJoinTable hashtable = hjstate->hj_HashTable;
2192 433524 : HashJoinTuple hashTuple = hjstate->hj_CurTuple;
2193 :
2194 : for (;;)
2195 : {
2196 : /*
2197 : * hj_CurTuple is the address of the tuple last returned from the
2198 : * current bucket, or NULL if it's time to start scanning a new
2199 : * bucket.
2200 : */
2201 5590872 : if (hashTuple != NULL)
2202 429594 : hashTuple = hashTuple->next.unshared;
2203 5161278 : else if (hjstate->hj_CurBucketNo < hashtable->nbuckets)
2204 : {
2205 5157360 : hashTuple = hashtable->buckets.unshared[hjstate->hj_CurBucketNo];
2206 5157360 : hjstate->hj_CurBucketNo++;
2207 : }
2208 3918 : else if (hjstate->hj_CurSkewBucketNo < hashtable->nSkewBuckets)
2209 : {
2210 0 : int j = hashtable->skewBucketNums[hjstate->hj_CurSkewBucketNo];
2211 :
2212 0 : hashTuple = hashtable->skewBucket[j]->tuples;
2213 0 : hjstate->hj_CurSkewBucketNo++;
2214 : }
2215 : else
2216 3918 : break; /* finished all buckets */
2217 :
2218 5998524 : while (hashTuple != NULL)
2219 : {
2220 841176 : if (!HeapTupleHeaderHasMatch(HJTUPLE_MINTUPLE(hashTuple)))
2221 : {
2222 : TupleTableSlot *inntuple;
2223 :
2224 : /* insert hashtable's tuple into exec slot */
2225 429606 : inntuple = ExecStoreMinimalTuple(HJTUPLE_MINTUPLE(hashTuple),
2226 : hjstate->hj_HashTupleSlot,
2227 : false); /* do not pfree */
2228 429606 : econtext->ecxt_innertuple = inntuple;
2229 :
2230 : /*
2231 : * Reset temp memory each time; although this function doesn't
2232 : * do any qual eval, the caller will, so let's keep it
2233 : * parallel to ExecScanHashBucket.
2234 : */
2235 429606 : ResetExprContext(econtext);
2236 :
2237 429606 : hjstate->hj_CurTuple = hashTuple;
2238 429606 : return true;
2239 : }
2240 :
2241 411570 : hashTuple = hashTuple->next.unshared;
2242 : }
2243 :
2244 : /* allow this loop to be cancellable */
2245 5157348 : CHECK_FOR_INTERRUPTS();
2246 : }
2247 :
2248 : /*
2249 : * no more unmatched tuples
2250 : */
2251 3918 : return false;
2252 : }
2253 :
2254 : /*
2255 : * ExecParallelScanHashTableForUnmatched
2256 : * scan the hash table for unmatched inner tuples, in parallel join
2257 : *
2258 : * On success, the inner tuple is stored into hjstate->hj_CurTuple and
2259 : * econtext->ecxt_innertuple, using hjstate->hj_HashTupleSlot as the slot
2260 : * for the latter.
2261 : */
2262 : bool
2263 120072 : ExecParallelScanHashTableForUnmatched(HashJoinState *hjstate,
2264 : ExprContext *econtext)
2265 : {
2266 120072 : HashJoinTable hashtable = hjstate->hj_HashTable;
2267 120072 : HashJoinTuple hashTuple = hjstate->hj_CurTuple;
2268 :
2269 : for (;;)
2270 : {
2271 : /*
2272 : * hj_CurTuple is the address of the tuple last returned from the
2273 : * current bucket, or NULL if it's time to start scanning a new
2274 : * bucket.
2275 : */
2276 734472 : if (hashTuple != NULL)
2277 120006 : hashTuple = ExecParallelHashNextTuple(hashtable, hashTuple);
2278 614466 : else if (hjstate->hj_CurBucketNo < hashtable->nbuckets)
2279 614400 : hashTuple = ExecParallelHashFirstTuple(hashtable,
2280 614400 : hjstate->hj_CurBucketNo++);
2281 : else
2282 66 : break; /* finished all buckets */
2283 :
2284 974406 : while (hashTuple != NULL)
2285 : {
2286 360006 : if (!HeapTupleHeaderHasMatch(HJTUPLE_MINTUPLE(hashTuple)))
2287 : {
2288 : TupleTableSlot *inntuple;
2289 :
2290 : /* insert hashtable's tuple into exec slot */
2291 120006 : inntuple = ExecStoreMinimalTuple(HJTUPLE_MINTUPLE(hashTuple),
2292 : hjstate->hj_HashTupleSlot,
2293 : false); /* do not pfree */
2294 120006 : econtext->ecxt_innertuple = inntuple;
2295 :
2296 : /*
2297 : * Reset temp memory each time; although this function doesn't
2298 : * do any qual eval, the caller will, so let's keep it
2299 : * parallel to ExecScanHashBucket.
2300 : */
2301 120006 : ResetExprContext(econtext);
2302 :
2303 120006 : hjstate->hj_CurTuple = hashTuple;
2304 120006 : return true;
2305 : }
2306 :
2307 240000 : hashTuple = ExecParallelHashNextTuple(hashtable, hashTuple);
2308 : }
2309 :
2310 : /* allow this loop to be cancellable */
2311 614400 : CHECK_FOR_INTERRUPTS();
2312 : }
2313 :
2314 : /*
2315 : * no more unmatched tuples
2316 : */
2317 66 : return false;
2318 : }
2319 :
2320 : /*
2321 : * ExecHashTableReset
2322 : *
2323 : * reset hash table header for new batch
2324 : */
2325 : void
2326 952 : ExecHashTableReset(HashJoinTable hashtable)
2327 : {
2328 : MemoryContext oldcxt;
2329 952 : int nbuckets = hashtable->nbuckets;
2330 :
2331 : /*
2332 : * Release all the hash buckets and tuples acquired in the prior pass, and
2333 : * reinitialize the context for a new pass.
2334 : */
2335 952 : MemoryContextReset(hashtable->batchCxt);
2336 952 : oldcxt = MemoryContextSwitchTo(hashtable->batchCxt);
2337 :
2338 : /* Reallocate and reinitialize the hash bucket headers. */
2339 952 : hashtable->buckets.unshared = palloc0_array(HashJoinTuple, nbuckets);
2340 :
2341 952 : hashtable->spaceUsed = 0;
2342 :
2343 952 : MemoryContextSwitchTo(oldcxt);
2344 :
2345 : /* Forget the chunks (the memory was freed by the context reset above). */
2346 952 : hashtable->chunks = NULL;
2347 952 : }
2348 :
2349 : /*
2350 : * ExecHashTableResetMatchFlags
2351 : * Clear all the HeapTupleHeaderHasMatch flags in the table
2352 : */
2353 : void
2354 70 : ExecHashTableResetMatchFlags(HashJoinTable hashtable)
2355 : {
2356 : HashJoinTuple tuple;
2357 : int i;
2358 :
2359 : /* Reset all flags in the main table ... */
2360 71750 : for (i = 0; i < hashtable->nbuckets; i++)
2361 : {
2362 72014 : for (tuple = hashtable->buckets.unshared[i]; tuple != NULL;
2363 334 : tuple = tuple->next.unshared)
2364 334 : HeapTupleHeaderClearMatch(HJTUPLE_MINTUPLE(tuple));
2365 : }
2366 :
2367 : /* ... and the same for the skew buckets, if any */
2368 70 : for (i = 0; i < hashtable->nSkewBuckets; i++)
2369 : {
2370 0 : int j = hashtable->skewBucketNums[i];
2371 0 : HashSkewBucket *skewBucket = hashtable->skewBucket[j];
2372 :
2373 0 : for (tuple = skewBucket->tuples; tuple != NULL; tuple = tuple->next.unshared)
2374 0 : HeapTupleHeaderClearMatch(HJTUPLE_MINTUPLE(tuple));
2375 : }
2376 70 : }
2377 :
2378 :
2379 : void
2380 1824 : ExecReScanHash(HashState *node)
2381 : {
2382 1824 : PlanState *outerPlan = outerPlanState(node);
2383 :
2384 : /*
2385 : * if chgParam of subnode is not null then plan will be re-scanned by
2386 : * first ExecProcNode.
2387 : */
2388 1824 : if (outerPlan->chgParam == NULL)
2389 30 : ExecReScan(outerPlan);
2390 1824 : }
2391 :
2392 :
2393 : /*
2394 : * ExecHashBuildSkewHash
2395 : *
2396 : * Set up for skew optimization if we can identify the most common values
2397 : * (MCVs) of the outer relation's join key. We make a skew hash bucket
2398 : * for the hash value of each MCV, up to the number of slots allowed
2399 : * based on available memory.
2400 : */
2401 : static void
2402 126 : ExecHashBuildSkewHash(HashState *hashstate, HashJoinTable hashtable,
2403 : Hash *node, int mcvsToUse)
2404 : {
2405 : HeapTupleData *statsTuple;
2406 : AttStatsSlot sslot;
2407 :
2408 : /* Do nothing if planner didn't identify the outer relation's join key */
2409 126 : if (!OidIsValid(node->skewTable))
2410 72 : return;
2411 : /* Also, do nothing if we don't have room for at least one skew bucket */
2412 126 : if (mcvsToUse <= 0)
2413 0 : return;
2414 :
2415 : /*
2416 : * Try to find the MCV statistics for the outer relation's join key.
2417 : */
2418 126 : statsTuple = SearchSysCache3(STATRELATTINH,
2419 : ObjectIdGetDatum(node->skewTable),
2420 126 : Int16GetDatum(node->skewColumn),
2421 126 : BoolGetDatum(node->skewInherit));
2422 126 : if (!HeapTupleIsValid(statsTuple))
2423 72 : return;
2424 :
2425 54 : if (get_attstatsslot(&sslot, statsTuple,
2426 : STATISTIC_KIND_MCV, InvalidOid,
2427 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
2428 : {
2429 : double frac;
2430 : int nbuckets;
2431 : int i;
2432 :
2433 6 : if (mcvsToUse > sslot.nvalues)
2434 0 : mcvsToUse = sslot.nvalues;
2435 :
2436 : /*
2437 : * Calculate the expected fraction of outer relation that will
2438 : * participate in the skew optimization. If this isn't at least
2439 : * SKEW_MIN_OUTER_FRACTION, don't use skew optimization.
2440 : */
2441 6 : frac = 0;
2442 132 : for (i = 0; i < mcvsToUse; i++)
2443 126 : frac += sslot.numbers[i];
2444 6 : if (frac < SKEW_MIN_OUTER_FRACTION)
2445 : {
2446 0 : free_attstatsslot(&sslot);
2447 0 : ReleaseSysCache(statsTuple);
2448 0 : return;
2449 : }
2450 :
2451 : /*
2452 : * Okay, set up the skew hashtable.
2453 : *
2454 : * skewBucket[] is an open addressing hashtable with a power of 2 size
2455 : * that is greater than the number of MCV values. (This ensures there
2456 : * will be at least one null entry, so searches will always
2457 : * terminate.)
2458 : *
2459 : * Note: this code could fail if mcvsToUse exceeds INT_MAX/8 or
2460 : * MaxAllocSize/sizeof(void *)/8, but that is not currently possible
2461 : * since we limit pg_statistic entries to much less than that.
2462 : */
2463 6 : nbuckets = pg_nextpower2_32(mcvsToUse + 1);
2464 : /* use two more bits just to help avoid collisions */
2465 6 : nbuckets <<= 2;
2466 :
2467 6 : hashtable->skewEnabled = true;
2468 6 : hashtable->skewBucketLen = nbuckets;
2469 :
2470 : /*
2471 : * We allocate the bucket memory in the hashtable's batch context. It
2472 : * is only needed during the first batch, and this ensures it will be
2473 : * automatically removed once the first batch is done.
2474 : */
2475 6 : hashtable->skewBucket = (HashSkewBucket **)
2476 6 : MemoryContextAllocZero(hashtable->batchCxt,
2477 : nbuckets * sizeof(HashSkewBucket *));
2478 6 : hashtable->skewBucketNums = (int *)
2479 6 : MemoryContextAllocZero(hashtable->batchCxt,
2480 : mcvsToUse * sizeof(int));
2481 :
2482 6 : hashtable->spaceUsed += nbuckets * sizeof(HashSkewBucket *)
2483 6 : + mcvsToUse * sizeof(int);
2484 6 : hashtable->spaceUsedSkew += nbuckets * sizeof(HashSkewBucket *)
2485 6 : + mcvsToUse * sizeof(int);
2486 6 : if (hashtable->spaceUsed > hashtable->spacePeak)
2487 6 : hashtable->spacePeak = hashtable->spaceUsed;
2488 :
2489 : /*
2490 : * Create a skew bucket for each MCV hash value.
2491 : *
2492 : * Note: it is very important that we create the buckets in order of
2493 : * decreasing MCV frequency. If we have to remove some buckets, they
2494 : * must be removed in reverse order of creation (see notes in
2495 : * ExecHashRemoveNextSkewBucket) and we want the least common MCVs to
2496 : * be removed first.
2497 : */
2498 :
2499 132 : for (i = 0; i < mcvsToUse; i++)
2500 : {
2501 : uint32 hashvalue;
2502 : int bucket;
2503 :
2504 126 : hashvalue = DatumGetUInt32(FunctionCall1Coll(hashstate->skew_hashfunction,
2505 : hashstate->skew_collation,
2506 126 : sslot.values[i]));
2507 :
2508 : /*
2509 : * While we have not hit a hole in the hashtable and have not hit
2510 : * the desired bucket, we have collided with some previous hash
2511 : * value, so try the next bucket location. NB: this code must
2512 : * match ExecHashGetSkewBucket.
2513 : */
2514 126 : bucket = hashvalue & (nbuckets - 1);
2515 126 : while (hashtable->skewBucket[bucket] != NULL &&
2516 0 : hashtable->skewBucket[bucket]->hashvalue != hashvalue)
2517 0 : bucket = (bucket + 1) & (nbuckets - 1);
2518 :
2519 : /*
2520 : * If we found an existing bucket with the same hashvalue, leave
2521 : * it alone. It's okay for two MCVs to share a hashvalue.
2522 : */
2523 126 : if (hashtable->skewBucket[bucket] != NULL)
2524 0 : continue;
2525 :
2526 : /* Okay, create a new skew bucket for this hashvalue. */
2527 252 : hashtable->skewBucket[bucket] = (HashSkewBucket *)
2528 126 : MemoryContextAlloc(hashtable->batchCxt,
2529 : sizeof(HashSkewBucket));
2530 126 : hashtable->skewBucket[bucket]->hashvalue = hashvalue;
2531 126 : hashtable->skewBucket[bucket]->tuples = NULL;
2532 126 : hashtable->skewBucketNums[hashtable->nSkewBuckets] = bucket;
2533 126 : hashtable->nSkewBuckets++;
2534 126 : hashtable->spaceUsed += SKEW_BUCKET_OVERHEAD;
2535 126 : hashtable->spaceUsedSkew += SKEW_BUCKET_OVERHEAD;
2536 126 : if (hashtable->spaceUsed > hashtable->spacePeak)
2537 126 : hashtable->spacePeak = hashtable->spaceUsed;
2538 : }
2539 :
2540 6 : free_attstatsslot(&sslot);
2541 : }
2542 :
2543 54 : ReleaseSysCache(statsTuple);
2544 : }
2545 :
2546 : /*
2547 : * ExecHashGetSkewBucket
2548 : *
2549 : * Returns the index of the skew bucket for this hashvalue,
2550 : * or INVALID_SKEW_BUCKET_NO if the hashvalue is not
2551 : * associated with any active skew bucket.
2552 : */
2553 : int
2554 29936568 : ExecHashGetSkewBucket(HashJoinTable hashtable, uint32 hashvalue)
2555 : {
2556 : int bucket;
2557 :
2558 : /*
2559 : * Always return INVALID_SKEW_BUCKET_NO if not doing skew optimization (in
2560 : * particular, this happens after the initial batch is done).
2561 : */
2562 29936568 : if (!hashtable->skewEnabled)
2563 29816568 : return INVALID_SKEW_BUCKET_NO;
2564 :
2565 : /*
2566 : * Since skewBucketLen is a power of 2, we can do a modulo by ANDing.
2567 : */
2568 120000 : bucket = hashvalue & (hashtable->skewBucketLen - 1);
2569 :
2570 : /*
2571 : * While we have not hit a hole in the hashtable and have not hit the
2572 : * desired bucket, we have collided with some other hash value, so try the
2573 : * next bucket location.
2574 : */
2575 127830 : while (hashtable->skewBucket[bucket] != NULL &&
2576 10818 : hashtable->skewBucket[bucket]->hashvalue != hashvalue)
2577 7830 : bucket = (bucket + 1) & (hashtable->skewBucketLen - 1);
2578 :
2579 : /*
2580 : * Found the desired bucket?
2581 : */
2582 120000 : if (hashtable->skewBucket[bucket] != NULL)
2583 2988 : return bucket;
2584 :
2585 : /*
2586 : * There must not be any hashtable entry for this hash value.
2587 : */
2588 117012 : return INVALID_SKEW_BUCKET_NO;
2589 : }
2590 :
2591 : /*
2592 : * ExecHashSkewTableInsert
2593 : *
2594 : * Insert a tuple into the skew hashtable.
2595 : *
2596 : * This should generally match up with the current-batch case in
2597 : * ExecHashTableInsert.
2598 : */
2599 : static void
2600 588 : ExecHashSkewTableInsert(HashJoinTable hashtable,
2601 : TupleTableSlot *slot,
2602 : uint32 hashvalue,
2603 : int bucketNumber)
2604 : {
2605 : bool shouldFree;
2606 588 : MinimalTuple tuple = ExecFetchSlotMinimalTuple(slot, &shouldFree);
2607 : HashJoinTuple hashTuple;
2608 : int hashTupleSize;
2609 :
2610 : /* Create the HashJoinTuple */
2611 588 : hashTupleSize = HJTUPLE_OVERHEAD + tuple->t_len;
2612 588 : hashTuple = (HashJoinTuple) MemoryContextAlloc(hashtable->batchCxt,
2613 : hashTupleSize);
2614 588 : hashTuple->hashvalue = hashvalue;
2615 588 : memcpy(HJTUPLE_MINTUPLE(hashTuple), tuple, tuple->t_len);
2616 588 : HeapTupleHeaderClearMatch(HJTUPLE_MINTUPLE(hashTuple));
2617 :
2618 : /* Push it onto the front of the skew bucket's list */
2619 588 : hashTuple->next.unshared = hashtable->skewBucket[bucketNumber]->tuples;
2620 588 : hashtable->skewBucket[bucketNumber]->tuples = hashTuple;
2621 : Assert(hashTuple != hashTuple->next.unshared);
2622 :
2623 : /* Account for space used, and back off if we've used too much */
2624 588 : hashtable->spaceUsed += hashTupleSize;
2625 588 : hashtable->spaceUsedSkew += hashTupleSize;
2626 588 : if (hashtable->spaceUsed > hashtable->spacePeak)
2627 432 : hashtable->spacePeak = hashtable->spaceUsed;
2628 690 : while (hashtable->spaceUsedSkew > hashtable->spaceAllowedSkew)
2629 102 : ExecHashRemoveNextSkewBucket(hashtable);
2630 :
2631 : /* Check we are not over the total spaceAllowed, either */
2632 588 : if (hashtable->spaceUsed > hashtable->spaceAllowed)
2633 0 : ExecHashIncreaseNumBatches(hashtable);
2634 :
2635 588 : if (shouldFree)
2636 588 : heap_free_minimal_tuple(tuple);
2637 588 : }
2638 :
2639 : /*
2640 : * ExecHashRemoveNextSkewBucket
2641 : *
2642 : * Remove the least valuable skew bucket by pushing its tuples into
2643 : * the main hash table.
2644 : */
2645 : static void
2646 102 : ExecHashRemoveNextSkewBucket(HashJoinTable hashtable)
2647 : {
2648 : int bucketToRemove;
2649 : HashSkewBucket *bucket;
2650 : uint32 hashvalue;
2651 : int bucketno;
2652 : int batchno;
2653 : HashJoinTuple hashTuple;
2654 :
2655 : /* Locate the bucket to remove */
2656 102 : bucketToRemove = hashtable->skewBucketNums[hashtable->nSkewBuckets - 1];
2657 102 : bucket = hashtable->skewBucket[bucketToRemove];
2658 :
2659 : /*
2660 : * Calculate which bucket and batch the tuples belong to in the main
2661 : * hashtable. They all have the same hash value, so it's the same for all
2662 : * of them. Also note that it's not possible for nbatch to increase while
2663 : * we are processing the tuples.
2664 : */
2665 102 : hashvalue = bucket->hashvalue;
2666 102 : ExecHashGetBucketAndBatch(hashtable, hashvalue, &bucketno, &batchno);
2667 :
2668 : /* Process all tuples in the bucket */
2669 102 : hashTuple = bucket->tuples;
2670 450 : while (hashTuple != NULL)
2671 : {
2672 348 : HashJoinTuple nextHashTuple = hashTuple->next.unshared;
2673 : MinimalTuple tuple;
2674 : Size tupleSize;
2675 :
2676 : /*
2677 : * This code must agree with ExecHashTableInsert. We do not use
2678 : * ExecHashTableInsert directly as ExecHashTableInsert expects a
2679 : * TupleTableSlot while we already have HashJoinTuples.
2680 : */
2681 348 : tuple = HJTUPLE_MINTUPLE(hashTuple);
2682 348 : tupleSize = HJTUPLE_OVERHEAD + tuple->t_len;
2683 :
2684 : /* Decide whether to put the tuple in the hash table or a temp file */
2685 348 : if (batchno == hashtable->curbatch)
2686 : {
2687 : /* Move the tuple to the main hash table */
2688 : HashJoinTuple copyTuple;
2689 :
2690 : /*
2691 : * We must copy the tuple into the dense storage, else it will not
2692 : * be found by, eg, ExecHashIncreaseNumBatches.
2693 : */
2694 138 : copyTuple = (HashJoinTuple) dense_alloc(hashtable, tupleSize);
2695 138 : memcpy(copyTuple, hashTuple, tupleSize);
2696 138 : pfree(hashTuple);
2697 :
2698 138 : copyTuple->next.unshared = hashtable->buckets.unshared[bucketno];
2699 138 : hashtable->buckets.unshared[bucketno] = copyTuple;
2700 :
2701 : /* We have reduced skew space, but overall space doesn't change */
2702 138 : hashtable->spaceUsedSkew -= tupleSize;
2703 : }
2704 : else
2705 : {
2706 : /* Put the tuple into a temp file for later batches */
2707 : Assert(batchno > hashtable->curbatch);
2708 210 : ExecHashJoinSaveTuple(tuple, hashvalue,
2709 210 : &hashtable->innerBatchFile[batchno],
2710 : hashtable);
2711 210 : pfree(hashTuple);
2712 210 : hashtable->spaceUsed -= tupleSize;
2713 210 : hashtable->spaceUsedSkew -= tupleSize;
2714 : }
2715 :
2716 348 : hashTuple = nextHashTuple;
2717 :
2718 : /* allow this loop to be cancellable */
2719 348 : CHECK_FOR_INTERRUPTS();
2720 : }
2721 :
2722 : /*
2723 : * Free the bucket struct itself and reset the hashtable entry to NULL.
2724 : *
2725 : * NOTE: this is not nearly as simple as it looks on the surface, because
2726 : * of the possibility of collisions in the hashtable. Suppose that hash
2727 : * values A and B collide at a particular hashtable entry, and that A was
2728 : * entered first so B gets shifted to a different table entry. If we were
2729 : * to remove A first then ExecHashGetSkewBucket would mistakenly start
2730 : * reporting that B is not in the hashtable, because it would hit the NULL
2731 : * before finding B. However, we always remove entries in the reverse
2732 : * order of creation, so this failure cannot happen.
2733 : */
2734 102 : hashtable->skewBucket[bucketToRemove] = NULL;
2735 102 : hashtable->nSkewBuckets--;
2736 102 : pfree(bucket);
2737 102 : hashtable->spaceUsed -= SKEW_BUCKET_OVERHEAD;
2738 102 : hashtable->spaceUsedSkew -= SKEW_BUCKET_OVERHEAD;
2739 :
2740 : /*
2741 : * If we have removed all skew buckets then give up on skew optimization.
2742 : * Release the arrays since they aren't useful any more.
2743 : */
2744 102 : if (hashtable->nSkewBuckets == 0)
2745 : {
2746 0 : hashtable->skewEnabled = false;
2747 0 : pfree(hashtable->skewBucket);
2748 0 : pfree(hashtable->skewBucketNums);
2749 0 : hashtable->skewBucket = NULL;
2750 0 : hashtable->skewBucketNums = NULL;
2751 0 : hashtable->spaceUsed -= hashtable->spaceUsedSkew;
2752 0 : hashtable->spaceUsedSkew = 0;
2753 : }
2754 102 : }
2755 :
2756 : /*
2757 : * Reserve space in the DSM segment for instrumentation data.
2758 : */
2759 : void
2760 198 : ExecHashEstimate(HashState *node, ParallelContext *pcxt)
2761 : {
2762 : size_t size;
2763 :
2764 : /* don't need this if not instrumenting or no workers */
2765 198 : if (!node->ps.instrument || pcxt->nworkers == 0)
2766 114 : return;
2767 :
2768 84 : size = mul_size(pcxt->nworkers, sizeof(HashInstrumentation));
2769 84 : size = add_size(size, offsetof(SharedHashInfo, hinstrument));
2770 84 : shm_toc_estimate_chunk(&pcxt->estimator, size);
2771 84 : shm_toc_estimate_keys(&pcxt->estimator, 1);
2772 : }
2773 :
2774 : /*
2775 : * Set up a space in the DSM for all workers to record instrumentation data
2776 : * about their hash table.
2777 : */
2778 : void
2779 198 : ExecHashInitializeDSM(HashState *node, ParallelContext *pcxt)
2780 : {
2781 : size_t size;
2782 :
2783 : /* don't need this if not instrumenting or no workers */
2784 198 : if (!node->ps.instrument || pcxt->nworkers == 0)
2785 114 : return;
2786 :
2787 84 : size = offsetof(SharedHashInfo, hinstrument) +
2788 84 : pcxt->nworkers * sizeof(HashInstrumentation);
2789 84 : node->shared_info = (SharedHashInfo *) shm_toc_allocate(pcxt->toc, size);
2790 :
2791 : /* Each per-worker area must start out as zeroes. */
2792 84 : memset(node->shared_info, 0, size);
2793 :
2794 84 : node->shared_info->num_workers = pcxt->nworkers;
2795 84 : shm_toc_insert(pcxt->toc, node->ps.plan->plan_node_id,
2796 84 : node->shared_info);
2797 : }
2798 :
2799 : /*
2800 : * Locate the DSM space for hash table instrumentation data that we'll write
2801 : * to at shutdown time.
2802 : */
2803 : void
2804 558 : ExecHashInitializeWorker(HashState *node, ParallelWorkerContext *pwcxt)
2805 : {
2806 : SharedHashInfo *shared_info;
2807 :
2808 : /* don't need this if not instrumenting */
2809 558 : if (!node->ps.instrument)
2810 306 : return;
2811 :
2812 : /*
2813 : * Find our entry in the shared area, and set up a pointer to it so that
2814 : * we'll accumulate stats there when shutting down or rebuilding the hash
2815 : * table.
2816 : */
2817 : shared_info = (SharedHashInfo *)
2818 252 : shm_toc_lookup(pwcxt->toc, node->ps.plan->plan_node_id, false);
2819 252 : node->hinstrument = &shared_info->hinstrument[ParallelWorkerNumber];
2820 : }
2821 :
2822 : /*
2823 : * Collect EXPLAIN stats if needed, saving them into DSM memory if
2824 : * ExecHashInitializeWorker was called, or local storage if not. In the
2825 : * parallel case, this must be done in ExecShutdownHash() rather than
2826 : * ExecEndHash() because the latter runs after we've detached from the DSM
2827 : * segment.
2828 : */
2829 : void
2830 30834 : ExecShutdownHash(HashState *node)
2831 : {
2832 : /* Allocate save space if EXPLAIN'ing and we didn't do so already */
2833 30834 : if (node->ps.instrument && !node->hinstrument)
2834 114 : node->hinstrument = palloc0_object(HashInstrumentation);
2835 : /* Now accumulate data for the current (final) hash table */
2836 30834 : if (node->hinstrument && node->hashtable)
2837 334 : ExecHashAccumInstrumentation(node->hinstrument, node->hashtable);
2838 30834 : }
2839 :
2840 : /*
2841 : * Retrieve instrumentation data from workers before the DSM segment is
2842 : * detached, so that EXPLAIN can access it.
2843 : */
2844 : void
2845 84 : ExecHashRetrieveInstrumentation(HashState *node)
2846 : {
2847 84 : SharedHashInfo *shared_info = node->shared_info;
2848 : size_t size;
2849 :
2850 84 : if (shared_info == NULL)
2851 0 : return;
2852 :
2853 : /* Replace node->shared_info with a copy in backend-local memory. */
2854 84 : size = offsetof(SharedHashInfo, hinstrument) +
2855 84 : shared_info->num_workers * sizeof(HashInstrumentation);
2856 84 : node->shared_info = palloc(size);
2857 84 : memcpy(node->shared_info, shared_info, size);
2858 : }
2859 :
2860 : /*
2861 : * Accumulate instrumentation data from 'hashtable' into an
2862 : * initially-zeroed HashInstrumentation struct.
2863 : *
2864 : * This is used to merge information across successive hash table instances
2865 : * within a single plan node. We take the maximum values of each interesting
2866 : * number. The largest nbuckets and largest nbatch values might have occurred
2867 : * in different instances, so there's some risk of confusion from reporting
2868 : * unrelated numbers; but there's a bigger risk of misdiagnosing a performance
2869 : * issue if we don't report the largest values. Similarly, we want to report
2870 : * the largest spacePeak regardless of whether it happened in the same
2871 : * instance as the largest nbuckets or nbatch. All the instances should have
2872 : * the same nbuckets_original and nbatch_original; but there's little value
2873 : * in depending on that here, so handle them the same way.
2874 : */
2875 : void
2876 334 : ExecHashAccumInstrumentation(HashInstrumentation *instrument,
2877 : HashJoinTable hashtable)
2878 : {
2879 334 : instrument->nbuckets = Max(instrument->nbuckets,
2880 : hashtable->nbuckets);
2881 334 : instrument->nbuckets_original = Max(instrument->nbuckets_original,
2882 : hashtable->nbuckets_original);
2883 334 : instrument->nbatch = Max(instrument->nbatch,
2884 : hashtable->nbatch);
2885 334 : instrument->nbatch_original = Max(instrument->nbatch_original,
2886 : hashtable->nbatch_original);
2887 334 : instrument->space_peak = Max(instrument->space_peak,
2888 : hashtable->spacePeak);
2889 334 : }
2890 :
2891 : /*
2892 : * Allocate 'size' bytes from the currently active HashMemoryChunk
2893 : */
2894 : static void *
2895 9438990 : dense_alloc(HashJoinTable hashtable, Size size)
2896 : {
2897 : HashMemoryChunk newChunk;
2898 : char *ptr;
2899 :
2900 : /* just in case the size is not already aligned properly */
2901 9438990 : size = MAXALIGN(size);
2902 :
2903 : /*
2904 : * If tuple size is larger than threshold, allocate a separate chunk.
2905 : */
2906 9438990 : if (size > HASH_CHUNK_THRESHOLD)
2907 : {
2908 : /* allocate new chunk and put it at the beginning of the list */
2909 0 : newChunk = (HashMemoryChunk) MemoryContextAlloc(hashtable->batchCxt,
2910 : HASH_CHUNK_HEADER_SIZE + size);
2911 0 : newChunk->maxlen = size;
2912 0 : newChunk->used = size;
2913 0 : newChunk->ntuples = 1;
2914 :
2915 : /*
2916 : * Add this chunk to the list after the first existing chunk, so that
2917 : * we don't lose the remaining space in the "current" chunk.
2918 : */
2919 0 : if (hashtable->chunks != NULL)
2920 : {
2921 0 : newChunk->next = hashtable->chunks->next;
2922 0 : hashtable->chunks->next.unshared = newChunk;
2923 : }
2924 : else
2925 : {
2926 0 : newChunk->next.unshared = hashtable->chunks;
2927 0 : hashtable->chunks = newChunk;
2928 : }
2929 :
2930 0 : return HASH_CHUNK_DATA(newChunk);
2931 : }
2932 :
2933 : /*
2934 : * See if we have enough space for it in the current chunk (if any). If
2935 : * not, allocate a fresh chunk.
2936 : */
2937 9438990 : if ((hashtable->chunks == NULL) ||
2938 9416130 : (hashtable->chunks->maxlen - hashtable->chunks->used) < size)
2939 : {
2940 : /* allocate new chunk and put it at the beginning of the list */
2941 34912 : newChunk = (HashMemoryChunk) MemoryContextAlloc(hashtable->batchCxt,
2942 : HASH_CHUNK_HEADER_SIZE + HASH_CHUNK_SIZE);
2943 :
2944 34912 : newChunk->maxlen = HASH_CHUNK_SIZE;
2945 34912 : newChunk->used = size;
2946 34912 : newChunk->ntuples = 1;
2947 :
2948 34912 : newChunk->next.unshared = hashtable->chunks;
2949 34912 : hashtable->chunks = newChunk;
2950 :
2951 34912 : return HASH_CHUNK_DATA(newChunk);
2952 : }
2953 :
2954 : /* There is enough space in the current chunk, let's add the tuple */
2955 9404078 : ptr = HASH_CHUNK_DATA(hashtable->chunks) + hashtable->chunks->used;
2956 9404078 : hashtable->chunks->used += size;
2957 9404078 : hashtable->chunks->ntuples += 1;
2958 :
2959 : /* return pointer to the start of the tuple memory */
2960 9404078 : return ptr;
2961 : }
2962 :
2963 : /*
2964 : * Allocate space for a tuple in shared dense storage. This is equivalent to
2965 : * dense_alloc but for Parallel Hash using shared memory.
2966 : *
2967 : * While loading a tuple into shared memory, we might run out of memory and
2968 : * decide to repartition, or determine that the load factor is too high and
2969 : * decide to expand the bucket array, or discover that another participant has
2970 : * commanded us to help do that. Return NULL if number of buckets or batches
2971 : * has changed, indicating that the caller must retry (considering the
2972 : * possibility that the tuple no longer belongs in the same batch).
2973 : */
2974 : static HashJoinTuple
2975 2390856 : ExecParallelHashTupleAlloc(HashJoinTable hashtable, size_t size,
2976 : dsa_pointer *shared)
2977 : {
2978 2390856 : ParallelHashJoinState *pstate = hashtable->parallel_state;
2979 : dsa_pointer chunk_shared;
2980 : HashMemoryChunk chunk;
2981 : Size chunk_size;
2982 : HashJoinTuple result;
2983 2390856 : int curbatch = hashtable->curbatch;
2984 :
2985 2390856 : size = MAXALIGN(size);
2986 :
2987 : /*
2988 : * Fast path: if there is enough space in this backend's current chunk,
2989 : * then we can allocate without any locking.
2990 : */
2991 2390856 : chunk = hashtable->current_chunk;
2992 2390856 : if (chunk != NULL &&
2993 2389988 : size <= HASH_CHUNK_THRESHOLD &&
2994 2389988 : chunk->maxlen - chunk->used >= size)
2995 : {
2996 :
2997 2387118 : chunk_shared = hashtable->current_chunk_shared;
2998 : Assert(chunk == dsa_get_address(hashtable->area, chunk_shared));
2999 2387118 : *shared = chunk_shared + HASH_CHUNK_HEADER_SIZE + chunk->used;
3000 2387118 : result = (HashJoinTuple) (HASH_CHUNK_DATA(chunk) + chunk->used);
3001 2387118 : chunk->used += size;
3002 :
3003 : Assert(chunk->used <= chunk->maxlen);
3004 : Assert(result == dsa_get_address(hashtable->area, *shared));
3005 :
3006 2387118 : return result;
3007 : }
3008 :
3009 : /* Slow path: try to allocate a new chunk. */
3010 3738 : LWLockAcquire(&pstate->lock, LW_EXCLUSIVE);
3011 :
3012 : /*
3013 : * Check if we need to help increase the number of buckets or batches.
3014 : */
3015 3738 : if (pstate->growth == PHJ_GROWTH_NEED_MORE_BATCHES ||
3016 3702 : pstate->growth == PHJ_GROWTH_NEED_MORE_BUCKETS)
3017 : {
3018 162 : ParallelHashGrowth growth = pstate->growth;
3019 :
3020 162 : hashtable->current_chunk = NULL;
3021 162 : LWLockRelease(&pstate->lock);
3022 :
3023 : /* Another participant has commanded us to help grow. */
3024 162 : if (growth == PHJ_GROWTH_NEED_MORE_BATCHES)
3025 36 : ExecParallelHashIncreaseNumBatches(hashtable);
3026 126 : else if (growth == PHJ_GROWTH_NEED_MORE_BUCKETS)
3027 126 : ExecParallelHashIncreaseNumBuckets(hashtable);
3028 :
3029 : /* The caller must retry. */
3030 162 : return NULL;
3031 : }
3032 :
3033 : /* Oversized tuples get their own chunk. */
3034 3576 : if (size > HASH_CHUNK_THRESHOLD)
3035 48 : chunk_size = size + HASH_CHUNK_HEADER_SIZE;
3036 : else
3037 3528 : chunk_size = HASH_CHUNK_SIZE;
3038 :
3039 : /* Check if it's time to grow batches or buckets. */
3040 3576 : if (pstate->growth != PHJ_GROWTH_DISABLED)
3041 : {
3042 : Assert(curbatch == 0);
3043 : Assert(BarrierPhase(&pstate->build_barrier) == PHJ_BUILD_HASH_INNER);
3044 :
3045 : /*
3046 : * Check if our space limit would be exceeded. To avoid choking on
3047 : * very large tuples or very low hash_mem setting, we'll always allow
3048 : * each backend to allocate at least one chunk.
3049 : */
3050 1886 : if (hashtable->batches[0].at_least_one_chunk &&
3051 1466 : hashtable->batches[0].shared->size +
3052 1466 : chunk_size > pstate->space_allowed)
3053 : {
3054 36 : pstate->growth = PHJ_GROWTH_NEED_MORE_BATCHES;
3055 36 : hashtable->batches[0].shared->space_exhausted = true;
3056 36 : LWLockRelease(&pstate->lock);
3057 :
3058 36 : return NULL;
3059 : }
3060 :
3061 : /* Check if our load factor limit would be exceeded. */
3062 1850 : if (hashtable->nbatch == 1)
3063 : {
3064 1574 : hashtable->batches[0].shared->ntuples += hashtable->batches[0].ntuples;
3065 1574 : hashtable->batches[0].ntuples = 0;
3066 : /* Guard against integer overflow and alloc size overflow */
3067 1574 : if (hashtable->batches[0].shared->ntuples + 1 >
3068 1574 : hashtable->nbuckets * NTUP_PER_BUCKET &&
3069 108 : hashtable->nbuckets < (INT_MAX / 2) &&
3070 108 : hashtable->nbuckets * 2 <=
3071 : MaxAllocSize / sizeof(dsa_pointer_atomic))
3072 : {
3073 108 : pstate->growth = PHJ_GROWTH_NEED_MORE_BUCKETS;
3074 108 : LWLockRelease(&pstate->lock);
3075 :
3076 108 : return NULL;
3077 : }
3078 : }
3079 : }
3080 :
3081 : /* We are cleared to allocate a new chunk. */
3082 3432 : chunk_shared = dsa_allocate(hashtable->area, chunk_size);
3083 3432 : hashtable->batches[curbatch].shared->size += chunk_size;
3084 3432 : hashtable->batches[curbatch].at_least_one_chunk = true;
3085 :
3086 : /* Set up the chunk. */
3087 3432 : chunk = (HashMemoryChunk) dsa_get_address(hashtable->area, chunk_shared);
3088 3432 : *shared = chunk_shared + HASH_CHUNK_HEADER_SIZE;
3089 3432 : chunk->maxlen = chunk_size - HASH_CHUNK_HEADER_SIZE;
3090 3432 : chunk->used = size;
3091 :
3092 : /*
3093 : * Push it onto the list of chunks, so that it can be found if we need to
3094 : * increase the number of buckets or batches (batch 0 only) and later for
3095 : * freeing the memory (all batches).
3096 : */
3097 3432 : chunk->next.shared = hashtable->batches[curbatch].shared->chunks;
3098 3432 : hashtable->batches[curbatch].shared->chunks = chunk_shared;
3099 :
3100 3432 : if (size <= HASH_CHUNK_THRESHOLD)
3101 : {
3102 : /*
3103 : * Make this the current chunk so that we can use the fast path to
3104 : * fill the rest of it up in future calls.
3105 : */
3106 3396 : hashtable->current_chunk = chunk;
3107 3396 : hashtable->current_chunk_shared = chunk_shared;
3108 : }
3109 3432 : LWLockRelease(&pstate->lock);
3110 :
3111 : Assert(HASH_CHUNK_DATA(chunk) == dsa_get_address(hashtable->area, *shared));
3112 3432 : result = (HashJoinTuple) HASH_CHUNK_DATA(chunk);
3113 :
3114 3432 : return result;
3115 : }
3116 :
3117 : /*
3118 : * One backend needs to set up the shared batch state including tuplestores.
3119 : * Other backends will ensure they have correctly configured accessors by
3120 : * called ExecParallelHashEnsureBatchAccessors().
3121 : */
3122 : static void
3123 226 : ExecParallelHashJoinSetUpBatches(HashJoinTable hashtable, int nbatch)
3124 : {
3125 226 : ParallelHashJoinState *pstate = hashtable->parallel_state;
3126 : ParallelHashJoinBatch *batches;
3127 : MemoryContext oldcxt;
3128 : int i;
3129 :
3130 : Assert(hashtable->batches == NULL);
3131 :
3132 : /* Allocate space. */
3133 226 : pstate->batches =
3134 226 : dsa_allocate0(hashtable->area,
3135 : EstimateParallelHashJoinBatch(hashtable) * nbatch);
3136 226 : pstate->nbatch = nbatch;
3137 226 : batches = dsa_get_address(hashtable->area, pstate->batches);
3138 :
3139 : /*
3140 : * Use hash join spill memory context to allocate accessors, including
3141 : * buffers for the temporary files.
3142 : */
3143 226 : oldcxt = MemoryContextSwitchTo(hashtable->spillCxt);
3144 :
3145 : /* Allocate this backend's accessor array. */
3146 226 : hashtable->nbatch = nbatch;
3147 226 : hashtable->batches =
3148 226 : palloc0_array(ParallelHashJoinBatchAccessor, hashtable->nbatch);
3149 :
3150 : /* Set up the shared state, tuplestores and backend-local accessors. */
3151 954 : for (i = 0; i < hashtable->nbatch; ++i)
3152 : {
3153 728 : ParallelHashJoinBatchAccessor *accessor = &hashtable->batches[i];
3154 728 : ParallelHashJoinBatch *shared = NthParallelHashJoinBatch(batches, i);
3155 : char name[MAXPGPATH];
3156 :
3157 : /*
3158 : * All members of shared were zero-initialized. We just need to set
3159 : * up the Barrier.
3160 : */
3161 728 : BarrierInit(&shared->batch_barrier, 0);
3162 728 : if (i == 0)
3163 : {
3164 : /* Batch 0 doesn't need to be loaded. */
3165 226 : BarrierAttach(&shared->batch_barrier);
3166 904 : while (BarrierPhase(&shared->batch_barrier) < PHJ_BATCH_PROBE)
3167 678 : BarrierArriveAndWait(&shared->batch_barrier, 0);
3168 226 : BarrierDetach(&shared->batch_barrier);
3169 : }
3170 :
3171 : /* Initialize accessor state. All members were zero-initialized. */
3172 728 : accessor->shared = shared;
3173 :
3174 : /* Initialize the shared tuplestores. */
3175 728 : snprintf(name, sizeof(name), "i%dof%d", i, hashtable->nbatch);
3176 728 : accessor->inner_tuples =
3177 728 : sts_initialize(ParallelHashJoinBatchInner(shared),
3178 : pstate->nparticipants,
3179 : ParallelWorkerNumber + 1,
3180 : sizeof(uint32),
3181 : SHARED_TUPLESTORE_SINGLE_PASS,
3182 : &pstate->fileset,
3183 : name);
3184 728 : snprintf(name, sizeof(name), "o%dof%d", i, hashtable->nbatch);
3185 728 : accessor->outer_tuples =
3186 728 : sts_initialize(ParallelHashJoinBatchOuter(shared,
3187 : pstate->nparticipants),
3188 : pstate->nparticipants,
3189 : ParallelWorkerNumber + 1,
3190 : sizeof(uint32),
3191 : SHARED_TUPLESTORE_SINGLE_PASS,
3192 : &pstate->fileset,
3193 : name);
3194 : }
3195 :
3196 226 : MemoryContextSwitchTo(oldcxt);
3197 226 : }
3198 :
3199 : /*
3200 : * Free the current set of ParallelHashJoinBatchAccessor objects.
3201 : */
3202 : static void
3203 56 : ExecParallelHashCloseBatchAccessors(HashJoinTable hashtable)
3204 : {
3205 : int i;
3206 :
3207 172 : for (i = 0; i < hashtable->nbatch; ++i)
3208 : {
3209 : /* Make sure no files are left open. */
3210 116 : sts_end_write(hashtable->batches[i].inner_tuples);
3211 116 : sts_end_write(hashtable->batches[i].outer_tuples);
3212 116 : sts_end_parallel_scan(hashtable->batches[i].inner_tuples);
3213 116 : sts_end_parallel_scan(hashtable->batches[i].outer_tuples);
3214 : }
3215 56 : pfree(hashtable->batches);
3216 56 : hashtable->batches = NULL;
3217 56 : }
3218 :
3219 : /*
3220 : * Make sure this backend has up-to-date accessors for the current set of
3221 : * batches.
3222 : */
3223 : static void
3224 936 : ExecParallelHashEnsureBatchAccessors(HashJoinTable hashtable)
3225 : {
3226 936 : ParallelHashJoinState *pstate = hashtable->parallel_state;
3227 : ParallelHashJoinBatch *batches;
3228 : MemoryContext oldcxt;
3229 : int i;
3230 :
3231 936 : if (hashtable->batches != NULL)
3232 : {
3233 694 : if (hashtable->nbatch == pstate->nbatch)
3234 692 : return;
3235 2 : ExecParallelHashCloseBatchAccessors(hashtable);
3236 : }
3237 :
3238 : /*
3239 : * We should never see a state where the batch-tracking array is freed,
3240 : * because we should have given up sooner if we join when the build
3241 : * barrier has reached the PHJ_BUILD_FREE phase.
3242 : */
3243 : Assert(DsaPointerIsValid(pstate->batches));
3244 :
3245 : /*
3246 : * Use hash join spill memory context to allocate accessors, including
3247 : * buffers for the temporary files.
3248 : */
3249 244 : oldcxt = MemoryContextSwitchTo(hashtable->spillCxt);
3250 :
3251 : /* Allocate this backend's accessor array. */
3252 244 : hashtable->nbatch = pstate->nbatch;
3253 244 : hashtable->batches =
3254 244 : palloc0_array(ParallelHashJoinBatchAccessor, hashtable->nbatch);
3255 :
3256 : /* Find the base of the pseudo-array of ParallelHashJoinBatch objects. */
3257 : batches = (ParallelHashJoinBatch *)
3258 244 : dsa_get_address(hashtable->area, pstate->batches);
3259 :
3260 : /* Set up the accessor array and attach to the tuplestores. */
3261 1164 : for (i = 0; i < hashtable->nbatch; ++i)
3262 : {
3263 920 : ParallelHashJoinBatchAccessor *accessor = &hashtable->batches[i];
3264 920 : ParallelHashJoinBatch *shared = NthParallelHashJoinBatch(batches, i);
3265 :
3266 920 : accessor->shared = shared;
3267 920 : accessor->preallocated = 0;
3268 920 : accessor->done = false;
3269 920 : accessor->outer_eof = false;
3270 920 : accessor->inner_tuples =
3271 920 : sts_attach(ParallelHashJoinBatchInner(shared),
3272 : ParallelWorkerNumber + 1,
3273 : &pstate->fileset);
3274 920 : accessor->outer_tuples =
3275 920 : sts_attach(ParallelHashJoinBatchOuter(shared,
3276 : pstate->nparticipants),
3277 : ParallelWorkerNumber + 1,
3278 : &pstate->fileset);
3279 : }
3280 :
3281 244 : MemoryContextSwitchTo(oldcxt);
3282 : }
3283 :
3284 : /*
3285 : * Allocate an empty shared memory hash table for a given batch.
3286 : */
3287 : void
3288 628 : ExecParallelHashTableAlloc(HashJoinTable hashtable, int batchno)
3289 : {
3290 628 : ParallelHashJoinBatch *batch = hashtable->batches[batchno].shared;
3291 : dsa_pointer_atomic *buckets;
3292 628 : int nbuckets = hashtable->parallel_state->nbuckets;
3293 : int i;
3294 :
3295 628 : batch->buckets =
3296 628 : dsa_allocate(hashtable->area, sizeof(dsa_pointer_atomic) * nbuckets);
3297 : buckets = (dsa_pointer_atomic *)
3298 628 : dsa_get_address(hashtable->area, batch->buckets);
3299 3187316 : for (i = 0; i < nbuckets; ++i)
3300 3186688 : dsa_pointer_atomic_init(&buckets[i], InvalidDsaPointer);
3301 628 : }
3302 :
3303 : /*
3304 : * If we are currently attached to a shared hash join batch, detach. If we
3305 : * are last to detach, clean up.
3306 : */
3307 : void
3308 24568 : ExecHashTableDetachBatch(HashJoinTable hashtable)
3309 : {
3310 24568 : if (hashtable->parallel_state != NULL &&
3311 1198 : hashtable->curbatch >= 0)
3312 : {
3313 784 : int curbatch = hashtable->curbatch;
3314 784 : ParallelHashJoinBatch *batch = hashtable->batches[curbatch].shared;
3315 784 : bool attached = true;
3316 :
3317 : /* Make sure any temporary files are closed. */
3318 784 : sts_end_parallel_scan(hashtable->batches[curbatch].inner_tuples);
3319 784 : sts_end_parallel_scan(hashtable->batches[curbatch].outer_tuples);
3320 :
3321 : /* After attaching we always get at least to PHJ_BATCH_PROBE. */
3322 : Assert(BarrierPhase(&batch->batch_barrier) == PHJ_BATCH_PROBE ||
3323 : BarrierPhase(&batch->batch_barrier) == PHJ_BATCH_SCAN);
3324 :
3325 : /*
3326 : * If we're abandoning the PHJ_BATCH_PROBE phase early without having
3327 : * reached the end of it, it means the plan doesn't want any more
3328 : * tuples, and it is happy to abandon any tuples buffered in this
3329 : * process's subplans. For correctness, we can't allow any process to
3330 : * execute the PHJ_BATCH_SCAN phase, because we will never have the
3331 : * complete set of match bits. Therefore we skip emitting unmatched
3332 : * tuples in all backends (if this is a full/right join), as if those
3333 : * tuples were all due to be emitted by this process and it has
3334 : * abandoned them too.
3335 : */
3336 784 : if (BarrierPhase(&batch->batch_barrier) == PHJ_BATCH_PROBE &&
3337 716 : !hashtable->batches[curbatch].outer_eof)
3338 : {
3339 : /*
3340 : * This flag may be written to by multiple backends during
3341 : * PHJ_BATCH_PROBE phase, but will only be read in PHJ_BATCH_SCAN
3342 : * phase so requires no extra locking.
3343 : */
3344 0 : batch->skip_unmatched = true;
3345 : }
3346 :
3347 : /*
3348 : * Even if we aren't doing a full/right outer join, we'll step through
3349 : * the PHJ_BATCH_SCAN phase just to maintain the invariant that
3350 : * freeing happens in PHJ_BATCH_FREE, but that'll be wait-free.
3351 : */
3352 784 : if (BarrierPhase(&batch->batch_barrier) == PHJ_BATCH_PROBE)
3353 716 : attached = BarrierArriveAndDetachExceptLast(&batch->batch_barrier);
3354 784 : if (attached && BarrierArriveAndDetach(&batch->batch_barrier))
3355 : {
3356 : /*
3357 : * We are not longer attached to the batch barrier, but we're the
3358 : * process that was chosen to free resources and it's safe to
3359 : * assert the current phase. The ParallelHashJoinBatch can't go
3360 : * away underneath us while we are attached to the build barrier,
3361 : * making this access safe.
3362 : */
3363 : Assert(BarrierPhase(&batch->batch_barrier) == PHJ_BATCH_FREE);
3364 :
3365 : /* Free shared chunks and buckets. */
3366 3758 : while (DsaPointerIsValid(batch->chunks))
3367 : {
3368 : HashMemoryChunk chunk =
3369 3130 : dsa_get_address(hashtable->area, batch->chunks);
3370 3130 : dsa_pointer next = chunk->next.shared;
3371 :
3372 3130 : dsa_free(hashtable->area, batch->chunks);
3373 3130 : batch->chunks = next;
3374 : }
3375 628 : if (DsaPointerIsValid(batch->buckets))
3376 : {
3377 628 : dsa_free(hashtable->area, batch->buckets);
3378 628 : batch->buckets = InvalidDsaPointer;
3379 : }
3380 : }
3381 :
3382 : /*
3383 : * Track the largest batch we've been attached to. Though each
3384 : * backend might see a different subset of batches, explain.c will
3385 : * scan the results from all backends to find the largest value.
3386 : */
3387 784 : hashtable->spacePeak =
3388 784 : Max(hashtable->spacePeak,
3389 : batch->size + sizeof(dsa_pointer_atomic) * hashtable->nbuckets);
3390 :
3391 : /* Remember that we are not attached to a batch. */
3392 784 : hashtable->curbatch = -1;
3393 : }
3394 24568 : }
3395 :
3396 : /*
3397 : * Detach from all shared resources. If we are last to detach, clean up.
3398 : */
3399 : void
3400 23784 : ExecHashTableDetach(HashJoinTable hashtable)
3401 : {
3402 23784 : ParallelHashJoinState *pstate = hashtable->parallel_state;
3403 :
3404 : /*
3405 : * If we're involved in a parallel query, we must either have gotten all
3406 : * the way to PHJ_BUILD_RUN, or joined too late and be in PHJ_BUILD_FREE.
3407 : */
3408 : Assert(!pstate ||
3409 : BarrierPhase(&pstate->build_barrier) >= PHJ_BUILD_RUN);
3410 :
3411 23784 : if (pstate && BarrierPhase(&pstate->build_barrier) == PHJ_BUILD_RUN)
3412 : {
3413 : int i;
3414 :
3415 : /* Make sure any temporary files are closed. */
3416 414 : if (hashtable->batches)
3417 : {
3418 1946 : for (i = 0; i < hashtable->nbatch; ++i)
3419 : {
3420 1532 : sts_end_write(hashtable->batches[i].inner_tuples);
3421 1532 : sts_end_write(hashtable->batches[i].outer_tuples);
3422 1532 : sts_end_parallel_scan(hashtable->batches[i].inner_tuples);
3423 1532 : sts_end_parallel_scan(hashtable->batches[i].outer_tuples);
3424 : }
3425 : }
3426 :
3427 : /* If we're last to detach, clean up shared memory. */
3428 414 : if (BarrierArriveAndDetach(&pstate->build_barrier))
3429 : {
3430 : /*
3431 : * Late joining processes will see this state and give up
3432 : * immediately.
3433 : */
3434 : Assert(BarrierPhase(&pstate->build_barrier) == PHJ_BUILD_FREE);
3435 :
3436 174 : if (DsaPointerIsValid(pstate->batches))
3437 : {
3438 174 : dsa_free(hashtable->area, pstate->batches);
3439 174 : pstate->batches = InvalidDsaPointer;
3440 : }
3441 : }
3442 : }
3443 23784 : hashtable->parallel_state = NULL;
3444 23784 : }
3445 :
3446 : /*
3447 : * Get the first tuple in a given bucket identified by number.
3448 : */
3449 : static inline HashJoinTuple
3450 2780430 : ExecParallelHashFirstTuple(HashJoinTable hashtable, int bucketno)
3451 : {
3452 : HashJoinTuple tuple;
3453 : dsa_pointer p;
3454 :
3455 : Assert(hashtable->parallel_state);
3456 2780430 : p = dsa_pointer_atomic_read(&hashtable->buckets.shared[bucketno]);
3457 2780430 : tuple = (HashJoinTuple) dsa_get_address(hashtable->area, p);
3458 :
3459 2780430 : return tuple;
3460 : }
3461 :
3462 : /*
3463 : * Get the next tuple in the same bucket as 'tuple'.
3464 : */
3465 : static inline HashJoinTuple
3466 3794594 : ExecParallelHashNextTuple(HashJoinTable hashtable, HashJoinTuple tuple)
3467 : {
3468 : HashJoinTuple next;
3469 :
3470 : Assert(hashtable->parallel_state);
3471 3794594 : next = (HashJoinTuple) dsa_get_address(hashtable->area, tuple->next.shared);
3472 :
3473 3794594 : return next;
3474 : }
3475 :
3476 : /*
3477 : * Insert a tuple at the front of a chain of tuples in DSA memory atomically.
3478 : */
3479 : static inline void
3480 2946936 : ExecParallelHashPushTuple(dsa_pointer_atomic *head,
3481 : HashJoinTuple tuple,
3482 : dsa_pointer tuple_shared)
3483 : {
3484 : for (;;)
3485 : {
3486 2959412 : tuple->next.shared = dsa_pointer_atomic_read(head);
3487 2959412 : if (dsa_pointer_atomic_compare_exchange(head,
3488 2959412 : &tuple->next.shared,
3489 : tuple_shared))
3490 2946936 : break;
3491 : }
3492 2946936 : }
3493 :
3494 : /*
3495 : * Prepare to work on a given batch.
3496 : */
3497 : void
3498 1804 : ExecParallelHashTableSetCurrentBatch(HashJoinTable hashtable, int batchno)
3499 : {
3500 : Assert(hashtable->batches[batchno].shared->buckets != InvalidDsaPointer);
3501 :
3502 1804 : hashtable->curbatch = batchno;
3503 1804 : hashtable->buckets.shared = (dsa_pointer_atomic *)
3504 1804 : dsa_get_address(hashtable->area,
3505 1804 : hashtable->batches[batchno].shared->buckets);
3506 1804 : hashtable->nbuckets = hashtable->parallel_state->nbuckets;
3507 1804 : hashtable->log2_nbuckets = pg_ceil_log2_32(hashtable->nbuckets);
3508 1804 : hashtable->current_chunk = NULL;
3509 1804 : hashtable->current_chunk_shared = InvalidDsaPointer;
3510 1804 : hashtable->batches[batchno].at_least_one_chunk = false;
3511 1804 : }
3512 :
3513 : /*
3514 : * Take the next available chunk from the queue of chunks being worked on in
3515 : * parallel. Return NULL if there are none left. Otherwise return a pointer
3516 : * to the chunk, and set *shared to the DSA pointer to the chunk.
3517 : */
3518 : static HashMemoryChunk
3519 1164 : ExecParallelHashPopChunkQueue(HashJoinTable hashtable, dsa_pointer *shared)
3520 : {
3521 1164 : ParallelHashJoinState *pstate = hashtable->parallel_state;
3522 : HashMemoryChunk chunk;
3523 :
3524 1164 : LWLockAcquire(&pstate->lock, LW_EXCLUSIVE);
3525 1164 : if (DsaPointerIsValid(pstate->chunk_work_queue))
3526 : {
3527 984 : *shared = pstate->chunk_work_queue;
3528 : chunk = (HashMemoryChunk)
3529 984 : dsa_get_address(hashtable->area, *shared);
3530 984 : pstate->chunk_work_queue = chunk->next.shared;
3531 : }
3532 : else
3533 180 : chunk = NULL;
3534 1164 : LWLockRelease(&pstate->lock);
3535 :
3536 1164 : return chunk;
3537 : }
3538 :
3539 : /*
3540 : * Increase the space preallocated in this backend for a given inner batch by
3541 : * at least a given amount. This allows us to track whether a given batch
3542 : * would fit in memory when loaded back in. Also increase the number of
3543 : * batches or buckets if required.
3544 : *
3545 : * This maintains a running estimation of how much space will be taken when we
3546 : * load the batch back into memory by simulating the way chunks will be handed
3547 : * out to workers. It's not perfectly accurate because the tuples will be
3548 : * packed into memory chunks differently by ExecParallelHashTupleAlloc(), but
3549 : * it should be pretty close. It tends to overestimate by a fraction of a
3550 : * chunk per worker since all workers gang up to preallocate during hashing,
3551 : * but workers tend to reload batches alone if there are enough to go around,
3552 : * leaving fewer partially filled chunks. This effect is bounded by
3553 : * nparticipants.
3554 : *
3555 : * Return false if the number of batches or buckets has changed, and the
3556 : * caller should reconsider which batch a given tuple now belongs in and call
3557 : * again.
3558 : */
3559 : static bool
3560 1516 : ExecParallelHashTuplePrealloc(HashJoinTable hashtable, int batchno, size_t size)
3561 : {
3562 1516 : ParallelHashJoinState *pstate = hashtable->parallel_state;
3563 1516 : ParallelHashJoinBatchAccessor *batch = &hashtable->batches[batchno];
3564 1516 : size_t want = Max(size, HASH_CHUNK_SIZE - HASH_CHUNK_HEADER_SIZE);
3565 :
3566 : Assert(batchno > 0);
3567 : Assert(batchno < hashtable->nbatch);
3568 : Assert(size == MAXALIGN(size));
3569 :
3570 1516 : LWLockAcquire(&pstate->lock, LW_EXCLUSIVE);
3571 :
3572 : /* Has another participant commanded us to help grow? */
3573 1516 : if (pstate->growth == PHJ_GROWTH_NEED_MORE_BATCHES ||
3574 1498 : pstate->growth == PHJ_GROWTH_NEED_MORE_BUCKETS)
3575 : {
3576 18 : ParallelHashGrowth growth = pstate->growth;
3577 :
3578 18 : LWLockRelease(&pstate->lock);
3579 18 : if (growth == PHJ_GROWTH_NEED_MORE_BATCHES)
3580 18 : ExecParallelHashIncreaseNumBatches(hashtable);
3581 0 : else if (growth == PHJ_GROWTH_NEED_MORE_BUCKETS)
3582 0 : ExecParallelHashIncreaseNumBuckets(hashtable);
3583 :
3584 18 : return false;
3585 : }
3586 :
3587 1498 : if (pstate->growth != PHJ_GROWTH_DISABLED &&
3588 1270 : batch->at_least_one_chunk &&
3589 640 : (batch->shared->estimated_size + want + HASH_CHUNK_HEADER_SIZE
3590 640 : > pstate->space_allowed))
3591 : {
3592 : /*
3593 : * We have determined that this batch would exceed the space budget if
3594 : * loaded into memory. Command all participants to help repartition.
3595 : */
3596 16 : batch->shared->space_exhausted = true;
3597 16 : pstate->growth = PHJ_GROWTH_NEED_MORE_BATCHES;
3598 16 : LWLockRelease(&pstate->lock);
3599 :
3600 16 : return false;
3601 : }
3602 :
3603 1482 : batch->at_least_one_chunk = true;
3604 1482 : batch->shared->estimated_size += want + HASH_CHUNK_HEADER_SIZE;
3605 1482 : batch->preallocated = want;
3606 1482 : LWLockRelease(&pstate->lock);
3607 :
3608 1482 : return true;
3609 : }
3610 :
3611 : /*
3612 : * Calculate the limit on how much memory can be used by Hash and similar
3613 : * plan types. This is work_mem times hash_mem_multiplier, and is
3614 : * expressed in bytes.
3615 : *
3616 : * Exported for use by the planner, as well as other hash-like executor
3617 : * nodes. This is a rather random place for this, but there is no better
3618 : * place.
3619 : */
3620 : size_t
3621 1739680 : get_hash_memory_limit(void)
3622 : {
3623 : double mem_limit;
3624 :
3625 : /* Do initial calculation in double arithmetic */
3626 1739680 : mem_limit = (double) work_mem * hash_mem_multiplier * 1024.0;
3627 :
3628 : /* Clamp in case it doesn't fit in size_t */
3629 1739680 : mem_limit = Min(mem_limit, (double) SIZE_MAX);
3630 :
3631 1739680 : return (size_t) mem_limit;
3632 : }
|