Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * nodeHashjoin.c
4 : * Routines to handle hash join nodes
5 : *
6 : * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/executor/nodeHashjoin.c
12 : *
13 : * HASH JOIN
14 : *
15 : * This is based on the "hybrid hash join" algorithm described shortly in the
16 : * following page
17 : *
18 : * https://en.wikipedia.org/wiki/Hash_join#Hybrid_hash_join
19 : *
20 : * and in detail in the referenced paper:
21 : *
22 : * "An Adaptive Hash Join Algorithm for Multiuser Environments"
23 : * Hansjörg Zeller; Jim Gray (1990). Proceedings of the 16th VLDB conference.
24 : * Brisbane: 186–197.
25 : *
26 : * If the inner side tuples of a hash join do not fit in memory, the hash join
27 : * can be executed in multiple batches.
28 : *
29 : * If the statistics on the inner side relation are accurate, planner chooses a
30 : * multi-batch strategy and estimates the number of batches.
31 : *
32 : * The query executor measures the real size of the hashtable and increases the
33 : * number of batches if the hashtable grows too large.
34 : *
35 : * The number of batches is always a power of two, so an increase in the number
36 : * of batches doubles it.
37 : *
38 : * Serial hash join measures batch size lazily -- waiting until it is loading a
39 : * batch to determine if it will fit in memory. While inserting tuples into the
40 : * hashtable, serial hash join will, if that tuple were to exceed work_mem,
41 : * dump out the hashtable and reassign them either to other batch files or the
42 : * current batch resident in the hashtable.
43 : *
44 : * Parallel hash join, on the other hand, completes all changes to the number
45 : * of batches during the build phase. If it increases the number of batches, it
46 : * dumps out all the tuples from all batches and reassigns them to entirely new
47 : * batch files. Then it checks every batch to ensure it will fit in the space
48 : * budget for the query.
49 : *
50 : * In both parallel and serial hash join, the executor currently makes a best
51 : * effort. If a particular batch will not fit in memory, it tries doubling the
52 : * number of batches. If after a batch increase, there is a batch which
53 : * retained all or none of its tuples, the executor disables growth in the
54 : * number of batches globally. After growth is disabled, all batches that would
55 : * have previously triggered an increase in the number of batches instead
56 : * exceed the space allowed.
57 : *
58 : * PARALLELISM
59 : *
60 : * Hash joins can participate in parallel query execution in several ways. A
61 : * parallel-oblivious hash join is one where the node is unaware that it is
62 : * part of a parallel plan. In this case, a copy of the inner plan is used to
63 : * build a copy of the hash table in every backend, and the outer plan could
64 : * either be built from a partial or complete path, so that the results of the
65 : * hash join are correspondingly either partial or complete. A parallel-aware
66 : * hash join is one that behaves differently, coordinating work between
67 : * backends, and appears as Parallel Hash Join in EXPLAIN output. A Parallel
68 : * Hash Join always appears with a Parallel Hash node.
69 : *
70 : * Parallel-aware hash joins use the same per-backend state machine to track
71 : * progress through the hash join algorithm as parallel-oblivious hash joins.
72 : * In a parallel-aware hash join, there is also a shared state machine that
73 : * co-operating backends use to synchronize their local state machines and
74 : * program counters. The shared state machine is managed with a Barrier IPC
75 : * primitive. When all attached participants arrive at a barrier, the phase
76 : * advances and all waiting participants are released.
77 : *
78 : * When a participant begins working on a parallel hash join, it must first
79 : * figure out how much progress has already been made, because participants
80 : * don't wait for each other to begin. For this reason there are switch
81 : * statements at key points in the code where we have to synchronize our local
82 : * state machine with the phase, and then jump to the correct part of the
83 : * algorithm so that we can get started.
84 : *
85 : * One barrier called build_barrier is used to coordinate the hashing phases.
86 : * The phase is represented by an integer which begins at zero and increments
87 : * one by one, but in the code it is referred to by symbolic names as follows.
88 : * An asterisk indicates a phase that is performed by a single arbitrarily
89 : * chosen process.
90 : *
91 : * PHJ_BUILD_ELECT -- initial state
92 : * PHJ_BUILD_ALLOCATE* -- one sets up the batches and table 0
93 : * PHJ_BUILD_HASH_INNER -- all hash the inner rel
94 : * PHJ_BUILD_HASH_OUTER -- (multi-batch only) all hash the outer
95 : * PHJ_BUILD_RUN -- building done, probing can begin
96 : * PHJ_BUILD_FREE* -- all work complete, one frees batches
97 : *
98 : * While in the phase PHJ_BUILD_HASH_INNER a separate pair of barriers may
99 : * be used repeatedly as required to coordinate expansions in the number of
100 : * batches or buckets. Their phases are as follows:
101 : *
102 : * PHJ_GROW_BATCHES_ELECT -- initial state
103 : * PHJ_GROW_BATCHES_REALLOCATE* -- one allocates new batches
104 : * PHJ_GROW_BATCHES_REPARTITION -- all repartition
105 : * PHJ_GROW_BATCHES_DECIDE* -- one detects skew and cleans up
106 : * PHJ_GROW_BATCHES_FINISH -- finished one growth cycle
107 : *
108 : * PHJ_GROW_BUCKETS_ELECT -- initial state
109 : * PHJ_GROW_BUCKETS_REALLOCATE* -- one allocates new buckets
110 : * PHJ_GROW_BUCKETS_REINSERT -- all insert tuples
111 : *
112 : * If the planner got the number of batches and buckets right, those won't be
113 : * necessary, but on the other hand we might finish up needing to expand the
114 : * buckets or batches multiple times while hashing the inner relation to stay
115 : * within our memory budget and load factor target. For that reason it's a
116 : * separate pair of barriers using circular phases.
117 : *
118 : * The PHJ_BUILD_HASH_OUTER phase is required only for multi-batch joins,
119 : * because we need to divide the outer relation into batches up front in order
120 : * to be able to process batches entirely independently. In contrast, the
121 : * parallel-oblivious algorithm simply throws tuples 'forward' to 'later'
122 : * batches whenever it encounters them while scanning and probing, which it
123 : * can do because it processes batches in serial order.
124 : *
125 : * Once PHJ_BUILD_RUN is reached, backends then split up and process
126 : * different batches, or gang up and work together on probing batches if there
127 : * aren't enough to go around. For each batch there is a separate barrier
128 : * with the following phases:
129 : *
130 : * PHJ_BATCH_ELECT -- initial state
131 : * PHJ_BATCH_ALLOCATE* -- one allocates buckets
132 : * PHJ_BATCH_LOAD -- all load the hash table from disk
133 : * PHJ_BATCH_PROBE -- all probe
134 : * PHJ_BATCH_SCAN* -- one does right/right-anti/full unmatched scan
135 : * PHJ_BATCH_FREE* -- one frees memory
136 : *
137 : * Batch 0 is a special case, because it starts out in phase
138 : * PHJ_BATCH_PROBE; populating batch 0's hash table is done during
139 : * PHJ_BUILD_HASH_INNER so we can skip loading.
140 : *
141 : * Initially we try to plan for a single-batch hash join using the combined
142 : * hash_mem of all participants to create a large shared hash table. If that
143 : * turns out either at planning or execution time to be impossible then we
144 : * fall back to regular hash_mem sized hash tables.
145 : *
146 : * To avoid deadlocks, we never wait for any barrier unless it is known that
147 : * all other backends attached to it are actively executing the node or have
148 : * finished. Practically, that means that we never emit a tuple while attached
149 : * to a barrier, unless the barrier has reached a phase that means that no
150 : * process will wait on it again. We emit tuples while attached to the build
151 : * barrier in phase PHJ_BUILD_RUN, and to a per-batch barrier in phase
152 : * PHJ_BATCH_PROBE. These are advanced to PHJ_BUILD_FREE and PHJ_BATCH_SCAN
153 : * respectively without waiting, using BarrierArriveAndDetach() and
154 : * BarrierArriveAndDetachExceptLast() respectively. The last to detach
155 : * receives a different return value so that it knows that it's safe to
156 : * clean up. Any straggler process that attaches after that phase is reached
157 : * will see that it's too late to participate or access the relevant shared
158 : * memory objects.
159 : *
160 : *-------------------------------------------------------------------------
161 : */
162 :
163 : #include "postgres.h"
164 :
165 : #include "access/htup_details.h"
166 : #include "access/parallel.h"
167 : #include "executor/executor.h"
168 : #include "executor/hashjoin.h"
169 : #include "executor/nodeHash.h"
170 : #include "executor/nodeHashjoin.h"
171 : #include "miscadmin.h"
172 : #include "utils/sharedtuplestore.h"
173 : #include "utils/wait_event.h"
174 :
175 :
176 : /*
177 : * States of the ExecHashJoin state machine
178 : */
179 : #define HJ_BUILD_HASHTABLE 1
180 : #define HJ_NEED_NEW_OUTER 2
181 : #define HJ_SCAN_BUCKET 3
182 : #define HJ_FILL_OUTER_TUPLE 4
183 : #define HJ_FILL_INNER_TUPLES 5
184 : #define HJ_NEED_NEW_BATCH 6
185 :
186 : /* Returns true if doing null-fill on outer relation */
187 : #define HJ_FILL_OUTER(hjstate) ((hjstate)->hj_NullInnerTupleSlot != NULL)
188 : /* Returns true if doing null-fill on inner relation */
189 : #define HJ_FILL_INNER(hjstate) ((hjstate)->hj_NullOuterTupleSlot != NULL)
190 :
191 : static TupleTableSlot *ExecHashJoinOuterGetTuple(PlanState *outerNode,
192 : HashJoinState *hjstate,
193 : uint32 *hashvalue);
194 : static TupleTableSlot *ExecParallelHashJoinOuterGetTuple(PlanState *outerNode,
195 : HashJoinState *hjstate,
196 : uint32 *hashvalue);
197 : static TupleTableSlot *ExecHashJoinGetSavedTuple(HashJoinState *hjstate,
198 : BufFile *file,
199 : uint32 *hashvalue,
200 : TupleTableSlot *tupleSlot);
201 : static bool ExecHashJoinNewBatch(HashJoinState *hjstate);
202 : static bool ExecParallelHashJoinNewBatch(HashJoinState *hjstate);
203 : static void ExecParallelHashJoinPartitionOuter(HashJoinState *hjstate);
204 :
205 :
206 : /* ----------------------------------------------------------------
207 : * ExecHashJoinImpl
208 : *
209 : * This function implements the Hybrid Hashjoin algorithm. It is marked
210 : * with an always-inline attribute so that ExecHashJoin() and
211 : * ExecParallelHashJoin() can inline it. Compilers that respect the
212 : * attribute should create versions specialized for parallel == true and
213 : * parallel == false with unnecessary branches removed.
214 : *
215 : * Note: the relation we build hash table on is the "inner"
216 : * the other one is "outer".
217 : * ----------------------------------------------------------------
218 : */
219 : static pg_attribute_always_inline TupleTableSlot *
220 9369912 : ExecHashJoinImpl(PlanState *pstate, bool parallel)
221 : {
222 9369912 : HashJoinState *node = castNode(HashJoinState, pstate);
223 : PlanState *outerNode;
224 : HashState *hashNode;
225 : ExprState *joinqual;
226 : ExprState *otherqual;
227 : ExprContext *econtext;
228 : HashJoinTable hashtable;
229 : TupleTableSlot *outerTupleSlot;
230 : uint32 hashvalue;
231 : int batchno;
232 : ParallelHashJoinState *parallel_state;
233 :
234 : /*
235 : * get information from HashJoin node
236 : */
237 9369912 : joinqual = node->js.joinqual;
238 9369912 : otherqual = node->js.ps.qual;
239 9369912 : hashNode = (HashState *) innerPlanState(node);
240 9369912 : outerNode = outerPlanState(node);
241 9369912 : hashtable = node->hj_HashTable;
242 9369912 : econtext = node->js.ps.ps_ExprContext;
243 9369912 : parallel_state = hashNode->parallel_state;
244 :
245 : /*
246 : * Reset per-tuple memory context to free any expression evaluation
247 : * storage allocated in the previous tuple cycle.
248 : */
249 9369912 : ResetExprContext(econtext);
250 :
251 : /*
252 : * run the hash join state machine
253 : */
254 : for (;;)
255 : {
256 : /*
257 : * It's possible to iterate this loop many times before returning a
258 : * tuple, in some pathological cases such as needing to move much of
259 : * the current batch to a later batch. So let's check for interrupts
260 : * each time through.
261 : */
262 35745084 : CHECK_FOR_INTERRUPTS();
263 :
264 35745084 : switch (node->hj_JoinState)
265 : {
266 24380 : case HJ_BUILD_HASHTABLE:
267 :
268 : /*
269 : * First time through: build hash table for inner relation.
270 : */
271 : Assert(hashtable == NULL);
272 :
273 : /*
274 : * If the outer relation is completely empty, and it's not
275 : * right/right-anti/full join, we can quit without building
276 : * the hash table. However, for an inner join it is only a
277 : * win to check this when the outer relation's startup cost is
278 : * less than the projected cost of building the hash table.
279 : * Otherwise it's best to build the hash table first and see
280 : * if the inner relation is empty. (When it's a left join, we
281 : * should always make this check, since we aren't going to be
282 : * able to skip the join on the strength of an empty inner
283 : * relation anyway.)
284 : *
285 : * If we are rescanning the join, we make use of information
286 : * gained on the previous scan: don't bother to try the
287 : * prefetch if the previous scan found the outer relation
288 : * nonempty. This is not 100% reliable since with new
289 : * parameters the outer relation might yield different
290 : * results, but it's a good heuristic.
291 : *
292 : * The only way to make the check is to try to fetch a tuple
293 : * from the outer plan node. If we succeed, we have to stash
294 : * it away for later consumption by ExecHashJoinOuterGetTuple.
295 : */
296 24380 : if (HJ_FILL_INNER(node))
297 : {
298 : /* no chance to not build the hash table */
299 5326 : node->hj_FirstOuterTupleSlot = NULL;
300 : }
301 19054 : else if (parallel)
302 : {
303 : /*
304 : * The empty-outer optimization is not implemented for
305 : * shared hash tables, because no one participant can
306 : * determine that there are no outer tuples, and it's not
307 : * yet clear that it's worth the synchronization overhead
308 : * of reaching consensus to figure that out. So we have
309 : * to build the hash table.
310 : */
311 326 : node->hj_FirstOuterTupleSlot = NULL;
312 : }
313 18728 : else if (HJ_FILL_OUTER(node) ||
314 13920 : (outerNode->plan->startup_cost < hashNode->ps.plan->total_cost &&
315 12962 : !node->hj_OuterNotEmpty))
316 : {
317 16354 : node->hj_FirstOuterTupleSlot = ExecProcNode(outerNode);
318 16354 : if (TupIsNull(node->hj_FirstOuterTupleSlot))
319 : {
320 4144 : node->hj_OuterNotEmpty = false;
321 4144 : return NULL;
322 : }
323 : else
324 12210 : node->hj_OuterNotEmpty = true;
325 : }
326 : else
327 2374 : node->hj_FirstOuterTupleSlot = NULL;
328 :
329 : /*
330 : * Create the hash table. If using Parallel Hash, then
331 : * whoever gets here first will create the hash table and any
332 : * later arrivals will merely attach to it.
333 : */
334 20236 : hashtable = ExecHashTableCreate(hashNode,
335 : node->hj_HashOperators,
336 : node->hj_Collations,
337 20236 : HJ_FILL_INNER(node));
338 20236 : node->hj_HashTable = hashtable;
339 :
340 : /*
341 : * Execute the Hash node, to build the hash table. If using
342 : * Parallel Hash, then we'll try to help hashing unless we
343 : * arrived too late.
344 : */
345 20236 : hashNode->hashtable = hashtable;
346 20236 : (void) MultiExecProcNode((PlanState *) hashNode);
347 :
348 : /*
349 : * If the inner relation is completely empty, and we're not
350 : * doing a left outer join, we can quit without scanning the
351 : * outer relation.
352 : */
353 20236 : if (hashtable->totalTuples == 0 && !HJ_FILL_OUTER(node))
354 : {
355 1674 : if (parallel)
356 : {
357 : /*
358 : * Advance the build barrier to PHJ_BUILD_RUN before
359 : * proceeding so we can negotiate resource cleanup.
360 : */
361 6 : Barrier *build_barrier = ¶llel_state->build_barrier;
362 :
363 8 : while (BarrierPhase(build_barrier) < PHJ_BUILD_RUN)
364 2 : BarrierArriveAndWait(build_barrier, 0);
365 : }
366 1674 : return NULL;
367 : }
368 :
369 : /*
370 : * need to remember whether nbatch has increased since we
371 : * began scanning the outer relation
372 : */
373 18562 : hashtable->nbatch_outstart = hashtable->nbatch;
374 :
375 : /*
376 : * Reset OuterNotEmpty for scan. (It's OK if we fetched a
377 : * tuple above, because ExecHashJoinOuterGetTuple will
378 : * immediately set it again.)
379 : */
380 18562 : node->hj_OuterNotEmpty = false;
381 :
382 18562 : if (parallel)
383 : {
384 : Barrier *build_barrier;
385 :
386 392 : build_barrier = ¶llel_state->build_barrier;
387 : Assert(BarrierPhase(build_barrier) == PHJ_BUILD_HASH_OUTER ||
388 : BarrierPhase(build_barrier) == PHJ_BUILD_RUN ||
389 : BarrierPhase(build_barrier) == PHJ_BUILD_FREE);
390 392 : if (BarrierPhase(build_barrier) == PHJ_BUILD_HASH_OUTER)
391 : {
392 : /*
393 : * If multi-batch, we need to hash the outer relation
394 : * up front.
395 : */
396 256 : if (hashtable->nbatch > 1)
397 142 : ExecParallelHashJoinPartitionOuter(node);
398 256 : BarrierArriveAndWait(build_barrier,
399 : WAIT_EVENT_HASH_BUILD_HASH_OUTER);
400 : }
401 136 : else if (BarrierPhase(build_barrier) == PHJ_BUILD_FREE)
402 : {
403 : /*
404 : * If we attached so late that the job is finished and
405 : * the batch state has been freed, we can return
406 : * immediately.
407 : */
408 2 : return NULL;
409 : }
410 :
411 : /* Each backend should now select a batch to work on. */
412 : Assert(BarrierPhase(build_barrier) == PHJ_BUILD_RUN);
413 390 : hashtable->curbatch = -1;
414 390 : node->hj_JoinState = HJ_NEED_NEW_BATCH;
415 :
416 390 : continue;
417 : }
418 : else
419 18170 : node->hj_JoinState = HJ_NEED_NEW_OUTER;
420 :
421 : /* FALL THRU */
422 :
423 16837462 : case HJ_NEED_NEW_OUTER:
424 :
425 : /*
426 : * We don't have an outer tuple, try to get the next one
427 : */
428 16837462 : if (parallel)
429 : outerTupleSlot =
430 2160900 : ExecParallelHashJoinOuterGetTuple(outerNode, node,
431 : &hashvalue);
432 : else
433 : outerTupleSlot =
434 14676562 : ExecHashJoinOuterGetTuple(outerNode, node, &hashvalue);
435 :
436 16837462 : if (TupIsNull(outerTupleSlot))
437 : {
438 : /* end of batch, or maybe whole join */
439 20782 : if (HJ_FILL_INNER(node))
440 : {
441 : /* set up to scan for unmatched inner tuples */
442 5086 : if (parallel)
443 : {
444 : /*
445 : * Only one process is currently allow to handle
446 : * each batch's unmatched tuples, in a parallel
447 : * join.
448 : */
449 74 : if (ExecParallelPrepHashTableForUnmatched(node))
450 66 : node->hj_JoinState = HJ_FILL_INNER_TUPLES;
451 : else
452 8 : node->hj_JoinState = HJ_NEED_NEW_BATCH;
453 : }
454 : else
455 : {
456 5012 : ExecPrepHashTableForUnmatched(node);
457 5012 : node->hj_JoinState = HJ_FILL_INNER_TUPLES;
458 : }
459 : }
460 : else
461 15696 : node->hj_JoinState = HJ_NEED_NEW_BATCH;
462 20782 : continue;
463 : }
464 :
465 16816680 : econtext->ecxt_outertuple = outerTupleSlot;
466 16816680 : node->hj_MatchedOuter = false;
467 :
468 : /*
469 : * Find the corresponding bucket for this tuple in the main
470 : * hash table or skew hash table.
471 : */
472 16816680 : node->hj_CurHashValue = hashvalue;
473 16816680 : ExecHashGetBucketAndBatch(hashtable, hashvalue,
474 : &node->hj_CurBucketNo, &batchno);
475 16816680 : node->hj_CurSkewBucketNo = ExecHashGetSkewBucket(hashtable,
476 : hashvalue);
477 16816680 : node->hj_CurTuple = NULL;
478 :
479 : /*
480 : * The tuple might not belong to the current batch (where
481 : * "current batch" includes the skew buckets if any).
482 : */
483 16816680 : if (batchno != hashtable->curbatch &&
484 1471392 : node->hj_CurSkewBucketNo == INVALID_SKEW_BUCKET_NO)
485 : {
486 : bool shouldFree;
487 1470192 : MinimalTuple mintuple = ExecFetchSlotMinimalTuple(outerTupleSlot,
488 : &shouldFree);
489 :
490 : /*
491 : * Need to postpone this outer tuple to a later batch.
492 : * Save it in the corresponding outer-batch file.
493 : */
494 : Assert(parallel_state == NULL);
495 : Assert(batchno > hashtable->curbatch);
496 1470192 : ExecHashJoinSaveTuple(mintuple, hashvalue,
497 1470192 : &hashtable->outerBatchFile[batchno],
498 : hashtable);
499 :
500 1470192 : if (shouldFree)
501 1470192 : heap_free_minimal_tuple(mintuple);
502 :
503 : /* Loop around, staying in HJ_NEED_NEW_OUTER state */
504 1470192 : continue;
505 : }
506 :
507 : /* OK, let's scan the bucket for matches */
508 15346488 : node->hj_JoinState = HJ_SCAN_BUCKET;
509 :
510 : /* FALL THRU */
511 :
512 21803588 : case HJ_SCAN_BUCKET:
513 :
514 : /*
515 : * Scan the selected hash bucket for matches to current outer
516 : */
517 21803588 : if (parallel)
518 : {
519 4200054 : if (!ExecParallelScanHashBucket(node, econtext))
520 : {
521 : /* out of matches; check for possible outer-join fill */
522 2160030 : node->hj_JoinState = HJ_FILL_OUTER_TUPLE;
523 2160030 : continue;
524 : }
525 : }
526 : else
527 : {
528 17603534 : if (!ExecScanHashBucket(node, econtext))
529 : {
530 : /* out of matches; check for possible outer-join fill */
531 9813198 : node->hj_JoinState = HJ_FILL_OUTER_TUPLE;
532 9813198 : continue;
533 : }
534 : }
535 :
536 : /*
537 : * We've got a match, but still need to test non-hashed quals.
538 : * ExecScanHashBucket already set up all the state needed to
539 : * call ExecQual.
540 : *
541 : * If we pass the qual, then save state for next call and have
542 : * ExecProject form the projection, store it in the tuple
543 : * table, and return the slot.
544 : *
545 : * Only the joinquals determine tuple match status, but all
546 : * quals must pass to actually return the tuple.
547 : */
548 9830360 : if (joinqual == NULL || ExecQual(joinqual, econtext))
549 : {
550 9677582 : node->hj_MatchedOuter = true;
551 :
552 :
553 : /*
554 : * This is really only needed if HJ_FILL_INNER(node), but
555 : * we'll avoid the branch and just set it always.
556 : */
557 9677582 : if (!HeapTupleHeaderHasMatch(HJTUPLE_MINTUPLE(node->hj_CurTuple)))
558 5765032 : HeapTupleHeaderSetMatch(HJTUPLE_MINTUPLE(node->hj_CurTuple));
559 :
560 : /* In an antijoin, we never return a matched tuple */
561 9677582 : if (node->js.jointype == JOIN_ANTI)
562 : {
563 1543570 : node->hj_JoinState = HJ_NEED_NEW_OUTER;
564 1543570 : continue;
565 : }
566 :
567 : /*
568 : * In a right-antijoin, we never return a matched tuple.
569 : * And we need to stay on the current outer tuple to
570 : * continue scanning the inner side for matches.
571 : */
572 8134012 : if (node->js.jointype == JOIN_RIGHT_ANTI)
573 24622 : continue;
574 :
575 : /*
576 : * If we only need to join to the first matching inner
577 : * tuple, then consider returning this one, but after that
578 : * continue with next outer tuple.
579 : */
580 8109390 : if (node->js.single_match)
581 1829594 : node->hj_JoinState = HJ_NEED_NEW_OUTER;
582 :
583 8109390 : if (otherqual == NULL || ExecQual(otherqual, econtext))
584 7925170 : return ExecProject(node->js.ps.ps_ProjInfo);
585 : else
586 184220 : InstrCountFiltered2(node, 1);
587 : }
588 : else
589 152778 : InstrCountFiltered1(node, 1);
590 336998 : break;
591 :
592 11973228 : case HJ_FILL_OUTER_TUPLE:
593 :
594 : /*
595 : * The current outer tuple has run out of matches, so check
596 : * whether to emit a dummy outer-join tuple. Whether we emit
597 : * one or not, the next state is NEED_NEW_OUTER.
598 : */
599 11973228 : node->hj_JoinState = HJ_NEED_NEW_OUTER;
600 :
601 11973228 : if (!node->hj_MatchedOuter &&
602 6998300 : HJ_FILL_OUTER(node))
603 : {
604 : /*
605 : * Generate a fake join tuple with nulls for the inner
606 : * tuple, and return it if it passes the non-join quals.
607 : */
608 2043266 : econtext->ecxt_innertuple = node->hj_NullInnerTupleSlot;
609 :
610 2043266 : if (otherqual == NULL || ExecQual(otherqual, econtext))
611 982022 : return ExecProject(node->js.ps.ps_ProjInfo);
612 : else
613 1061244 : InstrCountFiltered2(node, 1);
614 : }
615 10991206 : break;
616 :
617 449924 : case HJ_FILL_INNER_TUPLES:
618 :
619 : /*
620 : * We have finished a batch, but we are doing
621 : * right/right-anti/full join, so any unmatched inner tuples
622 : * in the hashtable have to be emitted before we continue to
623 : * the next batch.
624 : */
625 779776 : if (!(parallel ? ExecParallelScanHashTableForUnmatched(node, econtext)
626 329852 : : ExecScanHashTableForUnmatched(node, econtext)))
627 : {
628 : /* no more unmatched tuples */
629 5066 : node->hj_JoinState = HJ_NEED_NEW_BATCH;
630 5066 : continue;
631 : }
632 :
633 : /*
634 : * Generate a fake join tuple with nulls for the outer tuple,
635 : * and return it if it passes the non-join quals.
636 : */
637 444858 : econtext->ecxt_outertuple = node->hj_NullOuterTupleSlot;
638 :
639 444858 : if (otherqual == NULL || ExecQual(otherqual, econtext))
640 437762 : return ExecProject(node->js.ps.ps_ProjInfo);
641 : else
642 7096 : InstrCountFiltered2(node, 1);
643 7096 : break;
644 :
645 21160 : case HJ_NEED_NEW_BATCH:
646 :
647 : /*
648 : * Try to advance to next batch. Done if there are no more.
649 : */
650 21160 : if (parallel)
651 : {
652 1260 : if (!ExecParallelHashJoinNewBatch(node))
653 390 : return NULL; /* end of parallel-aware join */
654 : }
655 : else
656 : {
657 19900 : if (!ExecHashJoinNewBatch(node))
658 18748 : return NULL; /* end of parallel-oblivious join */
659 : }
660 2022 : node->hj_JoinState = HJ_NEED_NEW_OUTER;
661 2022 : break;
662 :
663 0 : default:
664 0 : elog(ERROR, "unrecognized hashjoin state: %d",
665 : (int) node->hj_JoinState);
666 : }
667 : }
668 : }
669 :
670 : /* ----------------------------------------------------------------
671 : * ExecHashJoin
672 : *
673 : * Parallel-oblivious version.
674 : * ----------------------------------------------------------------
675 : */
676 : static TupleTableSlot * /* return: a tuple or NULL */
677 7089478 : ExecHashJoin(PlanState *pstate)
678 : {
679 : /*
680 : * On sufficiently smart compilers this should be inlined with the
681 : * parallel-aware branches removed.
682 : */
683 7089478 : return ExecHashJoinImpl(pstate, false);
684 : }
685 :
686 : /* ----------------------------------------------------------------
687 : * ExecParallelHashJoin
688 : *
689 : * Parallel-aware version.
690 : * ----------------------------------------------------------------
691 : */
692 : static TupleTableSlot * /* return: a tuple or NULL */
693 2280434 : ExecParallelHashJoin(PlanState *pstate)
694 : {
695 : /*
696 : * On sufficiently smart compilers this should be inlined with the
697 : * parallel-oblivious branches removed.
698 : */
699 2280434 : return ExecHashJoinImpl(pstate, true);
700 : }
701 :
702 : /* ----------------------------------------------------------------
703 : * ExecInitHashJoin
704 : *
705 : * Init routine for HashJoin node.
706 : * ----------------------------------------------------------------
707 : */
708 : HashJoinState *
709 29714 : ExecInitHashJoin(HashJoin *node, EState *estate, int eflags)
710 : {
711 : HashJoinState *hjstate;
712 : Plan *outerNode;
713 : Hash *hashNode;
714 : TupleDesc outerDesc,
715 : innerDesc;
716 : const TupleTableSlotOps *ops;
717 :
718 : /* check for unsupported flags */
719 : Assert(!(eflags & (EXEC_FLAG_BACKWARD | EXEC_FLAG_MARK)));
720 :
721 : /*
722 : * create state structure
723 : */
724 29714 : hjstate = makeNode(HashJoinState);
725 29714 : hjstate->js.ps.plan = (Plan *) node;
726 29714 : hjstate->js.ps.state = estate;
727 :
728 : /*
729 : * See ExecHashJoinInitializeDSM() and ExecHashJoinInitializeWorker()
730 : * where this function may be replaced with a parallel version, if we
731 : * managed to launch a parallel query.
732 : */
733 29714 : hjstate->js.ps.ExecProcNode = ExecHashJoin;
734 29714 : hjstate->js.jointype = node->join.jointype;
735 :
736 : /*
737 : * Miscellaneous initialization
738 : *
739 : * create expression context for node
740 : */
741 29714 : ExecAssignExprContext(estate, &hjstate->js.ps);
742 :
743 : /*
744 : * initialize child nodes
745 : *
746 : * Note: we could suppress the REWIND flag for the inner input, which
747 : * would amount to betting that the hash will be a single batch. Not
748 : * clear if this would be a win or not.
749 : */
750 29714 : outerNode = outerPlan(node);
751 29714 : hashNode = (Hash *) innerPlan(node);
752 :
753 29714 : outerPlanState(hjstate) = ExecInitNode(outerNode, estate, eflags);
754 29714 : outerDesc = ExecGetResultType(outerPlanState(hjstate));
755 29714 : innerPlanState(hjstate) = ExecInitNode((Plan *) hashNode, estate, eflags);
756 29714 : innerDesc = ExecGetResultType(innerPlanState(hjstate));
757 :
758 : /*
759 : * Initialize result slot, type and projection.
760 : */
761 29714 : ExecInitResultTupleSlotTL(&hjstate->js.ps, &TTSOpsVirtual);
762 29714 : ExecAssignProjectionInfo(&hjstate->js.ps, NULL);
763 :
764 : /*
765 : * tuple table initialization
766 : */
767 29714 : ops = ExecGetResultSlotOps(outerPlanState(hjstate), NULL);
768 29714 : hjstate->hj_OuterTupleSlot = ExecInitExtraTupleSlot(estate, outerDesc,
769 : ops);
770 :
771 : /*
772 : * detect whether we need only consider the first matching inner tuple
773 : */
774 44638 : hjstate->js.single_match = (node->join.inner_unique ||
775 14924 : node->join.jointype == JOIN_SEMI);
776 :
777 : /* set up null tuples for outer joins, if needed */
778 29714 : switch (node->join.jointype)
779 : {
780 18050 : case JOIN_INNER:
781 : case JOIN_SEMI:
782 18050 : break;
783 5252 : case JOIN_LEFT:
784 : case JOIN_ANTI:
785 5252 : hjstate->hj_NullInnerTupleSlot =
786 5252 : ExecInitNullTupleSlot(estate, innerDesc, &TTSOpsVirtual);
787 5252 : break;
788 5376 : case JOIN_RIGHT:
789 : case JOIN_RIGHT_ANTI:
790 5376 : hjstate->hj_NullOuterTupleSlot =
791 5376 : ExecInitNullTupleSlot(estate, outerDesc, &TTSOpsVirtual);
792 5376 : break;
793 1036 : case JOIN_FULL:
794 1036 : hjstate->hj_NullOuterTupleSlot =
795 1036 : ExecInitNullTupleSlot(estate, outerDesc, &TTSOpsVirtual);
796 1036 : hjstate->hj_NullInnerTupleSlot =
797 1036 : ExecInitNullTupleSlot(estate, innerDesc, &TTSOpsVirtual);
798 1036 : break;
799 0 : default:
800 0 : elog(ERROR, "unrecognized join type: %d",
801 : (int) node->join.jointype);
802 : }
803 :
804 : /*
805 : * now for some voodoo. our temporary tuple slot is actually the result
806 : * tuple slot of the Hash node (which is our inner plan). we can do this
807 : * because Hash nodes don't return tuples via ExecProcNode() -- instead
808 : * the hash join node uses ExecScanHashBucket() to get at the contents of
809 : * the hash table. -cim 6/9/91
810 : */
811 : {
812 29714 : HashState *hashstate = (HashState *) innerPlanState(hjstate);
813 29714 : TupleTableSlot *slot = hashstate->ps.ps_ResultTupleSlot;
814 :
815 29714 : hjstate->hj_HashTupleSlot = slot;
816 : }
817 :
818 : /*
819 : * initialize child expressions
820 : */
821 29714 : hjstate->js.ps.qual =
822 29714 : ExecInitQual(node->join.plan.qual, (PlanState *) hjstate);
823 29714 : hjstate->js.joinqual =
824 29714 : ExecInitQual(node->join.joinqual, (PlanState *) hjstate);
825 29714 : hjstate->hashclauses =
826 29714 : ExecInitQual(node->hashclauses, (PlanState *) hjstate);
827 :
828 : /*
829 : * initialize hash-specific info
830 : */
831 29714 : hjstate->hj_HashTable = NULL;
832 29714 : hjstate->hj_FirstOuterTupleSlot = NULL;
833 :
834 29714 : hjstate->hj_CurHashValue = 0;
835 29714 : hjstate->hj_CurBucketNo = 0;
836 29714 : hjstate->hj_CurSkewBucketNo = INVALID_SKEW_BUCKET_NO;
837 29714 : hjstate->hj_CurTuple = NULL;
838 :
839 29714 : hjstate->hj_OuterHashKeys = ExecInitExprList(node->hashkeys,
840 : (PlanState *) hjstate);
841 29714 : hjstate->hj_HashOperators = node->hashoperators;
842 29714 : hjstate->hj_Collations = node->hashcollations;
843 :
844 29714 : hjstate->hj_JoinState = HJ_BUILD_HASHTABLE;
845 29714 : hjstate->hj_MatchedOuter = false;
846 29714 : hjstate->hj_OuterNotEmpty = false;
847 :
848 29714 : return hjstate;
849 : }
850 :
851 : /* ----------------------------------------------------------------
852 : * ExecEndHashJoin
853 : *
854 : * clean up routine for HashJoin node
855 : * ----------------------------------------------------------------
856 : */
857 : void
858 29612 : ExecEndHashJoin(HashJoinState *node)
859 : {
860 : /*
861 : * Free hash table
862 : */
863 29612 : if (node->hj_HashTable)
864 : {
865 18586 : ExecHashTableDestroy(node->hj_HashTable);
866 18586 : node->hj_HashTable = NULL;
867 : }
868 :
869 : /*
870 : * clean up subtrees
871 : */
872 29612 : ExecEndNode(outerPlanState(node));
873 29612 : ExecEndNode(innerPlanState(node));
874 29612 : }
875 :
876 : /*
877 : * ExecHashJoinOuterGetTuple
878 : *
879 : * get the next outer tuple for a parallel oblivious hashjoin: either by
880 : * executing the outer plan node in the first pass, or from the temp
881 : * files for the hashjoin batches.
882 : *
883 : * Returns a null slot if no more outer tuples (within the current batch).
884 : *
885 : * On success, the tuple's hash value is stored at *hashvalue --- this is
886 : * either originally computed, or re-read from the temp file.
887 : */
888 : static TupleTableSlot *
889 14676562 : ExecHashJoinOuterGetTuple(PlanState *outerNode,
890 : HashJoinState *hjstate,
891 : uint32 *hashvalue)
892 : {
893 14676562 : HashJoinTable hashtable = hjstate->hj_HashTable;
894 14676562 : int curbatch = hashtable->curbatch;
895 : TupleTableSlot *slot;
896 :
897 14676562 : if (curbatch == 0) /* if it is the first pass */
898 : {
899 : /*
900 : * Check to see if first outer tuple was already fetched by
901 : * ExecHashJoin() and not used yet.
902 : */
903 13205218 : slot = hjstate->hj_FirstOuterTupleSlot;
904 13205218 : if (!TupIsNull(slot))
905 11594 : hjstate->hj_FirstOuterTupleSlot = NULL;
906 : else
907 13193624 : slot = ExecProcNode(outerNode);
908 :
909 13206032 : while (!TupIsNull(slot))
910 : {
911 : /*
912 : * We have to compute the tuple's hash value.
913 : */
914 13187272 : ExprContext *econtext = hjstate->js.ps.ps_ExprContext;
915 :
916 13187272 : econtext->ecxt_outertuple = slot;
917 13187272 : if (ExecHashGetHashValue(hashtable, econtext,
918 : hjstate->hj_OuterHashKeys,
919 : true, /* outer tuple */
920 13187272 : HJ_FILL_OUTER(hjstate),
921 : hashvalue))
922 : {
923 : /* remember outer relation is not empty for possible rescan */
924 13186458 : hjstate->hj_OuterNotEmpty = true;
925 :
926 13186458 : return slot;
927 : }
928 :
929 : /*
930 : * That tuple couldn't match because of a NULL, so discard it and
931 : * continue with the next one.
932 : */
933 814 : slot = ExecProcNode(outerNode);
934 : }
935 : }
936 1471344 : else if (curbatch < hashtable->nbatch)
937 : {
938 1471344 : BufFile *file = hashtable->outerBatchFile[curbatch];
939 :
940 : /*
941 : * In outer-join cases, we could get here even though the batch file
942 : * is empty.
943 : */
944 1471344 : if (file == NULL)
945 0 : return NULL;
946 :
947 1471344 : slot = ExecHashJoinGetSavedTuple(hjstate,
948 : file,
949 : hashvalue,
950 : hjstate->hj_OuterTupleSlot);
951 1471344 : if (!TupIsNull(slot))
952 1470192 : return slot;
953 : }
954 :
955 : /* End of this batch */
956 19912 : return NULL;
957 : }
958 :
959 : /*
960 : * ExecHashJoinOuterGetTuple variant for the parallel case.
961 : */
962 : static TupleTableSlot *
963 2160900 : ExecParallelHashJoinOuterGetTuple(PlanState *outerNode,
964 : HashJoinState *hjstate,
965 : uint32 *hashvalue)
966 : {
967 2160900 : HashJoinTable hashtable = hjstate->hj_HashTable;
968 2160900 : int curbatch = hashtable->curbatch;
969 : TupleTableSlot *slot;
970 :
971 : /*
972 : * In the Parallel Hash case we only run the outer plan directly for
973 : * single-batch hash joins. Otherwise we have to go to batch files, even
974 : * for batch 0.
975 : */
976 2160900 : if (curbatch == 0 && hashtable->nbatch == 1)
977 : {
978 960134 : slot = ExecProcNode(outerNode);
979 :
980 960134 : while (!TupIsNull(slot))
981 : {
982 960006 : ExprContext *econtext = hjstate->js.ps.ps_ExprContext;
983 :
984 960006 : econtext->ecxt_outertuple = slot;
985 960006 : if (ExecHashGetHashValue(hashtable, econtext,
986 : hjstate->hj_OuterHashKeys,
987 : true, /* outer tuple */
988 960006 : HJ_FILL_OUTER(hjstate),
989 : hashvalue))
990 960006 : return slot;
991 :
992 : /*
993 : * That tuple couldn't match because of a NULL, so discard it and
994 : * continue with the next one.
995 : */
996 0 : slot = ExecProcNode(outerNode);
997 : }
998 : }
999 1200766 : else if (curbatch < hashtable->nbatch)
1000 : {
1001 : MinimalTuple tuple;
1002 :
1003 1200766 : tuple = sts_parallel_scan_next(hashtable->batches[curbatch].outer_tuples,
1004 : hashvalue);
1005 1200766 : if (tuple != NULL)
1006 : {
1007 1200024 : ExecForceStoreMinimalTuple(tuple,
1008 : hjstate->hj_OuterTupleSlot,
1009 : false);
1010 1200024 : slot = hjstate->hj_OuterTupleSlot;
1011 1200024 : return slot;
1012 : }
1013 : else
1014 742 : ExecClearTuple(hjstate->hj_OuterTupleSlot);
1015 : }
1016 :
1017 : /* End of this batch */
1018 870 : hashtable->batches[curbatch].outer_eof = true;
1019 :
1020 870 : return NULL;
1021 : }
1022 :
1023 : /*
1024 : * ExecHashJoinNewBatch
1025 : * switch to a new hashjoin batch
1026 : *
1027 : * Returns true if successful, false if there are no more batches.
1028 : */
1029 : static bool
1030 19900 : ExecHashJoinNewBatch(HashJoinState *hjstate)
1031 : {
1032 19900 : HashJoinTable hashtable = hjstate->hj_HashTable;
1033 : int nbatch;
1034 : int curbatch;
1035 : BufFile *innerFile;
1036 : TupleTableSlot *slot;
1037 : uint32 hashvalue;
1038 :
1039 19900 : nbatch = hashtable->nbatch;
1040 19900 : curbatch = hashtable->curbatch;
1041 :
1042 19900 : if (curbatch > 0)
1043 : {
1044 : /*
1045 : * We no longer need the previous outer batch file; close it right
1046 : * away to free disk space.
1047 : */
1048 1152 : if (hashtable->outerBatchFile[curbatch])
1049 1152 : BufFileClose(hashtable->outerBatchFile[curbatch]);
1050 1152 : hashtable->outerBatchFile[curbatch] = NULL;
1051 : }
1052 : else /* we just finished the first batch */
1053 : {
1054 : /*
1055 : * Reset some of the skew optimization state variables, since we no
1056 : * longer need to consider skew tuples after the first batch. The
1057 : * memory context reset we are about to do will release the skew
1058 : * hashtable itself.
1059 : */
1060 18748 : hashtable->skewEnabled = false;
1061 18748 : hashtable->skewBucket = NULL;
1062 18748 : hashtable->skewBucketNums = NULL;
1063 18748 : hashtable->nSkewBuckets = 0;
1064 18748 : hashtable->spaceUsedSkew = 0;
1065 : }
1066 :
1067 : /*
1068 : * We can always skip over any batches that are completely empty on both
1069 : * sides. We can sometimes skip over batches that are empty on only one
1070 : * side, but there are exceptions:
1071 : *
1072 : * 1. In a left/full outer join, we have to process outer batches even if
1073 : * the inner batch is empty. Similarly, in a right/right-anti/full outer
1074 : * join, we have to process inner batches even if the outer batch is
1075 : * empty.
1076 : *
1077 : * 2. If we have increased nbatch since the initial estimate, we have to
1078 : * scan inner batches since they might contain tuples that need to be
1079 : * reassigned to later inner batches.
1080 : *
1081 : * 3. Similarly, if we have increased nbatch since starting the outer
1082 : * scan, we have to rescan outer batches in case they contain tuples that
1083 : * need to be reassigned.
1084 : */
1085 19900 : curbatch++;
1086 19900 : while (curbatch < nbatch &&
1087 1152 : (hashtable->outerBatchFile[curbatch] == NULL ||
1088 1152 : hashtable->innerBatchFile[curbatch] == NULL))
1089 : {
1090 0 : if (hashtable->outerBatchFile[curbatch] &&
1091 0 : HJ_FILL_OUTER(hjstate))
1092 0 : break; /* must process due to rule 1 */
1093 0 : if (hashtable->innerBatchFile[curbatch] &&
1094 0 : HJ_FILL_INNER(hjstate))
1095 0 : break; /* must process due to rule 1 */
1096 0 : if (hashtable->innerBatchFile[curbatch] &&
1097 0 : nbatch != hashtable->nbatch_original)
1098 0 : break; /* must process due to rule 2 */
1099 0 : if (hashtable->outerBatchFile[curbatch] &&
1100 0 : nbatch != hashtable->nbatch_outstart)
1101 0 : break; /* must process due to rule 3 */
1102 : /* We can ignore this batch. */
1103 : /* Release associated temp files right away. */
1104 0 : if (hashtable->innerBatchFile[curbatch])
1105 0 : BufFileClose(hashtable->innerBatchFile[curbatch]);
1106 0 : hashtable->innerBatchFile[curbatch] = NULL;
1107 0 : if (hashtable->outerBatchFile[curbatch])
1108 0 : BufFileClose(hashtable->outerBatchFile[curbatch]);
1109 0 : hashtable->outerBatchFile[curbatch] = NULL;
1110 0 : curbatch++;
1111 : }
1112 :
1113 19900 : if (curbatch >= nbatch)
1114 18748 : return false; /* no more batches */
1115 :
1116 1152 : hashtable->curbatch = curbatch;
1117 :
1118 : /*
1119 : * Reload the hash table with the new inner batch (which could be empty)
1120 : */
1121 1152 : ExecHashTableReset(hashtable);
1122 :
1123 1152 : innerFile = hashtable->innerBatchFile[curbatch];
1124 :
1125 1152 : if (innerFile != NULL)
1126 : {
1127 1152 : if (BufFileSeek(innerFile, 0, 0, SEEK_SET))
1128 0 : ereport(ERROR,
1129 : (errcode_for_file_access(),
1130 : errmsg("could not rewind hash-join temporary file")));
1131 :
1132 2353850 : while ((slot = ExecHashJoinGetSavedTuple(hjstate,
1133 : innerFile,
1134 : &hashvalue,
1135 : hjstate->hj_HashTupleSlot)))
1136 : {
1137 : /*
1138 : * NOTE: some tuples may be sent to future batches. Also, it is
1139 : * possible for hashtable->nbatch to be increased here!
1140 : */
1141 2352698 : ExecHashTableInsert(hashtable, slot, hashvalue);
1142 : }
1143 :
1144 : /*
1145 : * after we build the hash table, the inner batch file is no longer
1146 : * needed
1147 : */
1148 1152 : BufFileClose(innerFile);
1149 1152 : hashtable->innerBatchFile[curbatch] = NULL;
1150 : }
1151 :
1152 : /*
1153 : * Rewind outer batch file (if present), so that we can start reading it.
1154 : */
1155 1152 : if (hashtable->outerBatchFile[curbatch] != NULL)
1156 : {
1157 1152 : if (BufFileSeek(hashtable->outerBatchFile[curbatch], 0, 0, SEEK_SET))
1158 0 : ereport(ERROR,
1159 : (errcode_for_file_access(),
1160 : errmsg("could not rewind hash-join temporary file")));
1161 : }
1162 :
1163 1152 : return true;
1164 : }
1165 :
1166 : /*
1167 : * Choose a batch to work on, and attach to it. Returns true if successful,
1168 : * false if there are no more batches.
1169 : */
1170 : static bool
1171 1260 : ExecParallelHashJoinNewBatch(HashJoinState *hjstate)
1172 : {
1173 1260 : HashJoinTable hashtable = hjstate->hj_HashTable;
1174 : int start_batchno;
1175 : int batchno;
1176 :
1177 : /*
1178 : * If we were already attached to a batch, remember not to bother checking
1179 : * it again, and detach from it (possibly freeing the hash table if we are
1180 : * last to detach).
1181 : */
1182 1260 : if (hashtable->curbatch >= 0)
1183 : {
1184 862 : hashtable->batches[hashtable->curbatch].done = true;
1185 862 : ExecHashTableDetachBatch(hashtable);
1186 : }
1187 :
1188 : /*
1189 : * Search for a batch that isn't done. We use an atomic counter to start
1190 : * our search at a different batch in every participant when there are
1191 : * more batches than participants.
1192 : */
1193 1260 : batchno = start_batchno =
1194 1260 : pg_atomic_fetch_add_u32(&hashtable->parallel_state->distributor, 1) %
1195 1260 : hashtable->nbatch;
1196 : do
1197 : {
1198 : uint32 hashvalue;
1199 : MinimalTuple tuple;
1200 : TupleTableSlot *slot;
1201 :
1202 3050 : if (!hashtable->batches[batchno].done)
1203 : {
1204 : SharedTuplestoreAccessor *inner_tuples;
1205 1764 : Barrier *batch_barrier =
1206 1764 : &hashtable->batches[batchno].shared->batch_barrier;
1207 :
1208 1764 : switch (BarrierAttach(batch_barrier))
1209 : {
1210 582 : case PHJ_BATCH_ELECT:
1211 :
1212 : /* One backend allocates the hash table. */
1213 582 : if (BarrierArriveAndWait(batch_barrier,
1214 : WAIT_EVENT_HASH_BATCH_ELECT))
1215 582 : ExecParallelHashTableAlloc(hashtable, batchno);
1216 : /* Fall through. */
1217 :
1218 : case PHJ_BATCH_ALLOCATE:
1219 : /* Wait for allocation to complete. */
1220 582 : BarrierArriveAndWait(batch_barrier,
1221 : WAIT_EVENT_HASH_BATCH_ALLOCATE);
1222 : /* Fall through. */
1223 :
1224 596 : case PHJ_BATCH_LOAD:
1225 : /* Start (or join in) loading tuples. */
1226 596 : ExecParallelHashTableSetCurrentBatch(hashtable, batchno);
1227 596 : inner_tuples = hashtable->batches[batchno].inner_tuples;
1228 596 : sts_begin_parallel_scan(inner_tuples);
1229 1083524 : while ((tuple = sts_parallel_scan_next(inner_tuples,
1230 : &hashvalue)))
1231 : {
1232 1082928 : ExecForceStoreMinimalTuple(tuple,
1233 : hjstate->hj_HashTupleSlot,
1234 : false);
1235 1082928 : slot = hjstate->hj_HashTupleSlot;
1236 1082928 : ExecParallelHashTableInsertCurrentBatch(hashtable, slot,
1237 : hashvalue);
1238 : }
1239 596 : sts_end_parallel_scan(inner_tuples);
1240 596 : BarrierArriveAndWait(batch_barrier,
1241 : WAIT_EVENT_HASH_BATCH_LOAD);
1242 : /* Fall through. */
1243 :
1244 870 : case PHJ_BATCH_PROBE:
1245 :
1246 : /*
1247 : * This batch is ready to probe. Return control to
1248 : * caller. We stay attached to batch_barrier so that the
1249 : * hash table stays alive until everyone's finished
1250 : * probing it, but no participant is allowed to wait at
1251 : * this barrier again (or else a deadlock could occur).
1252 : * All attached participants must eventually detach from
1253 : * the barrier and one worker must advance the phase so
1254 : * that the final phase is reached.
1255 : */
1256 870 : ExecParallelHashTableSetCurrentBatch(hashtable, batchno);
1257 870 : sts_begin_parallel_scan(hashtable->batches[batchno].outer_tuples);
1258 :
1259 870 : return true;
1260 0 : case PHJ_BATCH_SCAN:
1261 :
1262 : /*
1263 : * In principle, we could help scan for unmatched tuples,
1264 : * since that phase is already underway (the thing we
1265 : * can't do under current deadlock-avoidance rules is wait
1266 : * for others to arrive at PHJ_BATCH_SCAN, because
1267 : * PHJ_BATCH_PROBE emits tuples, but in this case we just
1268 : * got here without waiting). That is not yet done. For
1269 : * now, we just detach and go around again. We have to
1270 : * use ExecHashTableDetachBatch() because there's a small
1271 : * chance we'll be the last to detach, and then we're
1272 : * responsible for freeing memory.
1273 : */
1274 0 : ExecParallelHashTableSetCurrentBatch(hashtable, batchno);
1275 0 : hashtable->batches[batchno].done = true;
1276 0 : ExecHashTableDetachBatch(hashtable);
1277 0 : break;
1278 :
1279 894 : case PHJ_BATCH_FREE:
1280 :
1281 : /*
1282 : * Already done. Detach and go around again (if any
1283 : * remain).
1284 : */
1285 894 : BarrierDetach(batch_barrier);
1286 894 : hashtable->batches[batchno].done = true;
1287 894 : hashtable->curbatch = -1;
1288 894 : break;
1289 :
1290 0 : default:
1291 0 : elog(ERROR, "unexpected batch phase %d",
1292 : BarrierPhase(batch_barrier));
1293 : }
1294 : }
1295 2180 : batchno = (batchno + 1) % hashtable->nbatch;
1296 2180 : } while (batchno != start_batchno);
1297 :
1298 390 : return false;
1299 : }
1300 :
1301 : /*
1302 : * ExecHashJoinSaveTuple
1303 : * save a tuple to a batch file.
1304 : *
1305 : * The data recorded in the file for each tuple is its hash value,
1306 : * then the tuple in MinimalTuple format.
1307 : *
1308 : * fileptr points to a batch file in one of the hashtable arrays.
1309 : *
1310 : * The batch files (and their buffers) are allocated in the spill context
1311 : * created for the hashtable.
1312 : */
1313 : void
1314 3822890 : ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
1315 : BufFile **fileptr, HashJoinTable hashtable)
1316 : {
1317 3822890 : BufFile *file = *fileptr;
1318 :
1319 : /*
1320 : * The batch file is lazily created. If this is the first tuple written to
1321 : * this batch, the batch file is created and its buffer is allocated in
1322 : * the spillCxt context, NOT in the batchCxt.
1323 : *
1324 : * During the build phase, buffered files are created for inner batches.
1325 : * Each batch's buffered file is closed (and its buffer freed) after the
1326 : * batch is loaded into memory during the outer side scan. Therefore, it
1327 : * is necessary to allocate the batch file buffer in a memory context
1328 : * which outlives the batch itself.
1329 : *
1330 : * Also, we use spillCxt instead of hashCxt for a better accounting of the
1331 : * spilling memory consumption.
1332 : */
1333 3822890 : if (file == NULL)
1334 : {
1335 2304 : MemoryContext oldctx = MemoryContextSwitchTo(hashtable->spillCxt);
1336 :
1337 2304 : file = BufFileCreateTemp(false);
1338 2304 : *fileptr = file;
1339 :
1340 2304 : MemoryContextSwitchTo(oldctx);
1341 : }
1342 :
1343 3822890 : BufFileWrite(file, &hashvalue, sizeof(uint32));
1344 3822890 : BufFileWrite(file, tuple, tuple->t_len);
1345 3822890 : }
1346 :
1347 : /*
1348 : * ExecHashJoinGetSavedTuple
1349 : * read the next tuple from a batch file. Return NULL if no more.
1350 : *
1351 : * On success, *hashvalue is set to the tuple's hash value, and the tuple
1352 : * itself is stored in the given slot.
1353 : */
1354 : static TupleTableSlot *
1355 3825194 : ExecHashJoinGetSavedTuple(HashJoinState *hjstate,
1356 : BufFile *file,
1357 : uint32 *hashvalue,
1358 : TupleTableSlot *tupleSlot)
1359 : {
1360 : uint32 header[2];
1361 : size_t nread;
1362 : MinimalTuple tuple;
1363 :
1364 : /*
1365 : * We check for interrupts here because this is typically taken as an
1366 : * alternative code path to an ExecProcNode() call, which would include
1367 : * such a check.
1368 : */
1369 3825194 : CHECK_FOR_INTERRUPTS();
1370 :
1371 : /*
1372 : * Since both the hash value and the MinimalTuple length word are uint32,
1373 : * we can read them both in one BufFileRead() call without any type
1374 : * cheating.
1375 : */
1376 3825194 : nread = BufFileReadMaybeEOF(file, header, sizeof(header), true);
1377 3825194 : if (nread == 0) /* end of file */
1378 : {
1379 2304 : ExecClearTuple(tupleSlot);
1380 2304 : return NULL;
1381 : }
1382 3822890 : *hashvalue = header[0];
1383 3822890 : tuple = (MinimalTuple) palloc(header[1]);
1384 3822890 : tuple->t_len = header[1];
1385 3822890 : BufFileReadExact(file,
1386 : (char *) tuple + sizeof(uint32),
1387 3822890 : header[1] - sizeof(uint32));
1388 3822890 : ExecForceStoreMinimalTuple(tuple, tupleSlot, true);
1389 3822890 : return tupleSlot;
1390 : }
1391 :
1392 :
1393 : void
1394 2790 : ExecReScanHashJoin(HashJoinState *node)
1395 : {
1396 2790 : PlanState *outerPlan = outerPlanState(node);
1397 2790 : PlanState *innerPlan = innerPlanState(node);
1398 :
1399 : /*
1400 : * In a multi-batch join, we currently have to do rescans the hard way,
1401 : * primarily because batch temp files may have already been released. But
1402 : * if it's a single-batch join, and there is no parameter change for the
1403 : * inner subnode, then we can just re-use the existing hash table without
1404 : * rebuilding it.
1405 : */
1406 2790 : if (node->hj_HashTable != NULL)
1407 : {
1408 2350 : if (node->hj_HashTable->nbatch == 1 &&
1409 2350 : innerPlan->chgParam == NULL)
1410 : {
1411 : /*
1412 : * Okay to reuse the hash table; needn't rescan inner, either.
1413 : *
1414 : * However, if it's a right/right-anti/full join, we'd better
1415 : * reset the inner-tuple match flags contained in the table.
1416 : */
1417 802 : if (HJ_FILL_INNER(node))
1418 14 : ExecHashTableResetMatchFlags(node->hj_HashTable);
1419 :
1420 : /*
1421 : * Also, we need to reset our state about the emptiness of the
1422 : * outer relation, so that the new scan of the outer will update
1423 : * it correctly if it turns out to be empty this time. (There's no
1424 : * harm in clearing it now because ExecHashJoin won't need the
1425 : * info. In the other cases, where the hash table doesn't exist
1426 : * or we are destroying it, we leave this state alone because
1427 : * ExecHashJoin will need it the first time through.)
1428 : */
1429 802 : node->hj_OuterNotEmpty = false;
1430 :
1431 : /* ExecHashJoin can skip the BUILD_HASHTABLE step */
1432 802 : node->hj_JoinState = HJ_NEED_NEW_OUTER;
1433 : }
1434 : else
1435 : {
1436 : /* must destroy and rebuild hash table */
1437 1548 : HashState *hashNode = castNode(HashState, innerPlan);
1438 :
1439 : Assert(hashNode->hashtable == node->hj_HashTable);
1440 : /* accumulate stats from old hash table, if wanted */
1441 : /* (this should match ExecShutdownHash) */
1442 1548 : if (hashNode->ps.instrument && !hashNode->hinstrument)
1443 0 : hashNode->hinstrument = (HashInstrumentation *)
1444 0 : palloc0(sizeof(HashInstrumentation));
1445 1548 : if (hashNode->hinstrument)
1446 0 : ExecHashAccumInstrumentation(hashNode->hinstrument,
1447 : hashNode->hashtable);
1448 : /* for safety, be sure to clear child plan node's pointer too */
1449 1548 : hashNode->hashtable = NULL;
1450 :
1451 1548 : ExecHashTableDestroy(node->hj_HashTable);
1452 1548 : node->hj_HashTable = NULL;
1453 1548 : node->hj_JoinState = HJ_BUILD_HASHTABLE;
1454 :
1455 : /*
1456 : * if chgParam of subnode is not null then plan will be re-scanned
1457 : * by first ExecProcNode.
1458 : */
1459 1548 : if (innerPlan->chgParam == NULL)
1460 0 : ExecReScan(innerPlan);
1461 : }
1462 : }
1463 :
1464 : /* Always reset intra-tuple state */
1465 2790 : node->hj_CurHashValue = 0;
1466 2790 : node->hj_CurBucketNo = 0;
1467 2790 : node->hj_CurSkewBucketNo = INVALID_SKEW_BUCKET_NO;
1468 2790 : node->hj_CurTuple = NULL;
1469 :
1470 2790 : node->hj_MatchedOuter = false;
1471 2790 : node->hj_FirstOuterTupleSlot = NULL;
1472 :
1473 : /*
1474 : * if chgParam of subnode is not null then plan will be re-scanned by
1475 : * first ExecProcNode.
1476 : */
1477 2790 : if (outerPlan->chgParam == NULL)
1478 2050 : ExecReScan(outerPlan);
1479 2790 : }
1480 :
1481 : void
1482 26286 : ExecShutdownHashJoin(HashJoinState *node)
1483 : {
1484 26286 : if (node->hj_HashTable)
1485 : {
1486 : /*
1487 : * Detach from shared state before DSM memory goes away. This makes
1488 : * sure that we don't have any pointers into DSM memory by the time
1489 : * ExecEndHashJoin runs.
1490 : */
1491 18568 : ExecHashTableDetachBatch(node->hj_HashTable);
1492 18568 : ExecHashTableDetach(node->hj_HashTable);
1493 : }
1494 26286 : }
1495 :
1496 : static void
1497 142 : ExecParallelHashJoinPartitionOuter(HashJoinState *hjstate)
1498 : {
1499 142 : PlanState *outerState = outerPlanState(hjstate);
1500 142 : ExprContext *econtext = hjstate->js.ps.ps_ExprContext;
1501 142 : HashJoinTable hashtable = hjstate->hj_HashTable;
1502 : TupleTableSlot *slot;
1503 : uint32 hashvalue;
1504 : int i;
1505 :
1506 : Assert(hjstate->hj_FirstOuterTupleSlot == NULL);
1507 :
1508 : /* Execute outer plan, writing all tuples to shared tuplestores. */
1509 : for (;;)
1510 : {
1511 1200166 : slot = ExecProcNode(outerState);
1512 1200166 : if (TupIsNull(slot))
1513 : break;
1514 1200024 : econtext->ecxt_outertuple = slot;
1515 1200024 : if (ExecHashGetHashValue(hashtable, econtext,
1516 : hjstate->hj_OuterHashKeys,
1517 : true, /* outer tuple */
1518 1200024 : HJ_FILL_OUTER(hjstate),
1519 : &hashvalue))
1520 : {
1521 : int batchno;
1522 : int bucketno;
1523 : bool shouldFree;
1524 1200024 : MinimalTuple mintup = ExecFetchSlotMinimalTuple(slot, &shouldFree);
1525 :
1526 1200024 : ExecHashGetBucketAndBatch(hashtable, hashvalue, &bucketno,
1527 : &batchno);
1528 1200024 : sts_puttuple(hashtable->batches[batchno].outer_tuples,
1529 : &hashvalue, mintup);
1530 :
1531 1200024 : if (shouldFree)
1532 1200024 : heap_free_minimal_tuple(mintup);
1533 : }
1534 1200024 : CHECK_FOR_INTERRUPTS();
1535 : }
1536 :
1537 : /* Make sure all outer partitions are readable by any backend. */
1538 1206 : for (i = 0; i < hashtable->nbatch; ++i)
1539 1064 : sts_end_write(hashtable->batches[i].outer_tuples);
1540 142 : }
1541 :
1542 : void
1543 120 : ExecHashJoinEstimate(HashJoinState *state, ParallelContext *pcxt)
1544 : {
1545 120 : shm_toc_estimate_chunk(&pcxt->estimator, sizeof(ParallelHashJoinState));
1546 120 : shm_toc_estimate_keys(&pcxt->estimator, 1);
1547 120 : }
1548 :
1549 : void
1550 120 : ExecHashJoinInitializeDSM(HashJoinState *state, ParallelContext *pcxt)
1551 : {
1552 120 : int plan_node_id = state->js.ps.plan->plan_node_id;
1553 : HashState *hashNode;
1554 : ParallelHashJoinState *pstate;
1555 :
1556 : /*
1557 : * Disable shared hash table mode if we failed to create a real DSM
1558 : * segment, because that means that we don't have a DSA area to work with.
1559 : */
1560 120 : if (pcxt->seg == NULL)
1561 0 : return;
1562 :
1563 120 : ExecSetExecProcNode(&state->js.ps, ExecParallelHashJoin);
1564 :
1565 : /*
1566 : * Set up the state needed to coordinate access to the shared hash
1567 : * table(s), using the plan node ID as the toc key.
1568 : */
1569 120 : pstate = shm_toc_allocate(pcxt->toc, sizeof(ParallelHashJoinState));
1570 120 : shm_toc_insert(pcxt->toc, plan_node_id, pstate);
1571 :
1572 : /*
1573 : * Set up the shared hash join state with no batches initially.
1574 : * ExecHashTableCreate() will prepare at least one later and set nbatch
1575 : * and space_allowed.
1576 : */
1577 120 : pstate->nbatch = 0;
1578 120 : pstate->space_allowed = 0;
1579 120 : pstate->batches = InvalidDsaPointer;
1580 120 : pstate->old_batches = InvalidDsaPointer;
1581 120 : pstate->nbuckets = 0;
1582 120 : pstate->growth = PHJ_GROWTH_OK;
1583 120 : pstate->chunk_work_queue = InvalidDsaPointer;
1584 120 : pg_atomic_init_u32(&pstate->distributor, 0);
1585 120 : pstate->nparticipants = pcxt->nworkers + 1;
1586 120 : pstate->total_tuples = 0;
1587 120 : LWLockInitialize(&pstate->lock,
1588 : LWTRANCHE_PARALLEL_HASH_JOIN);
1589 120 : BarrierInit(&pstate->build_barrier, 0);
1590 120 : BarrierInit(&pstate->grow_batches_barrier, 0);
1591 120 : BarrierInit(&pstate->grow_buckets_barrier, 0);
1592 :
1593 : /* Set up the space we'll use for shared temporary files. */
1594 120 : SharedFileSetInit(&pstate->fileset, pcxt->seg);
1595 :
1596 : /* Initialize the shared state in the hash node. */
1597 120 : hashNode = (HashState *) innerPlanState(state);
1598 120 : hashNode->parallel_state = pstate;
1599 : }
1600 :
1601 : /* ----------------------------------------------------------------
1602 : * ExecHashJoinReInitializeDSM
1603 : *
1604 : * Reset shared state before beginning a fresh scan.
1605 : * ----------------------------------------------------------------
1606 : */
1607 : void
1608 48 : ExecHashJoinReInitializeDSM(HashJoinState *state, ParallelContext *pcxt)
1609 : {
1610 48 : int plan_node_id = state->js.ps.plan->plan_node_id;
1611 : ParallelHashJoinState *pstate =
1612 48 : shm_toc_lookup(pcxt->toc, plan_node_id, false);
1613 :
1614 : /*
1615 : * It would be possible to reuse the shared hash table in single-batch
1616 : * cases by resetting and then fast-forwarding build_barrier to
1617 : * PHJ_BUILD_FREE and batch 0's batch_barrier to PHJ_BATCH_PROBE, but
1618 : * currently shared hash tables are already freed by now (by the last
1619 : * participant to detach from the batch). We could consider keeping it
1620 : * around for single-batch joins. We'd also need to adjust
1621 : * finalize_plan() so that it doesn't record a dummy dependency for
1622 : * Parallel Hash nodes, preventing the rescan optimization. For now we
1623 : * don't try.
1624 : */
1625 :
1626 : /* Detach, freeing any remaining shared memory. */
1627 48 : if (state->hj_HashTable != NULL)
1628 : {
1629 0 : ExecHashTableDetachBatch(state->hj_HashTable);
1630 0 : ExecHashTableDetach(state->hj_HashTable);
1631 : }
1632 :
1633 : /* Clear any shared batch files. */
1634 48 : SharedFileSetDeleteAll(&pstate->fileset);
1635 :
1636 : /* Reset build_barrier to PHJ_BUILD_ELECT so we can go around again. */
1637 48 : BarrierInit(&pstate->build_barrier, 0);
1638 48 : }
1639 :
1640 : void
1641 308 : ExecHashJoinInitializeWorker(HashJoinState *state,
1642 : ParallelWorkerContext *pwcxt)
1643 : {
1644 : HashState *hashNode;
1645 308 : int plan_node_id = state->js.ps.plan->plan_node_id;
1646 : ParallelHashJoinState *pstate =
1647 308 : shm_toc_lookup(pwcxt->toc, plan_node_id, false);
1648 :
1649 : /* Attach to the space for shared temporary files. */
1650 308 : SharedFileSetAttach(&pstate->fileset, pwcxt->seg);
1651 :
1652 : /* Attach to the shared state in the hash node. */
1653 308 : hashNode = (HashState *) innerPlanState(state);
1654 308 : hashNode->parallel_state = pstate;
1655 :
1656 308 : ExecSetExecProcNode(&state->js.ps, ExecParallelHashJoin);
1657 308 : }
|