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/lsyscache.h"
173 : #include "utils/sharedtuplestore.h"
174 : #include "utils/wait_event.h"
175 :
176 :
177 : /*
178 : * States of the ExecHashJoin state machine
179 : */
180 : #define HJ_BUILD_HASHTABLE 1
181 : #define HJ_NEED_NEW_OUTER 2
182 : #define HJ_SCAN_BUCKET 3
183 : #define HJ_FILL_OUTER_TUPLE 4
184 : #define HJ_FILL_INNER_TUPLES 5
185 : #define HJ_NEED_NEW_BATCH 6
186 :
187 : /* Returns true if doing null-fill on outer relation */
188 : #define HJ_FILL_OUTER(hjstate) ((hjstate)->hj_NullInnerTupleSlot != NULL)
189 : /* Returns true if doing null-fill on inner relation */
190 : #define HJ_FILL_INNER(hjstate) ((hjstate)->hj_NullOuterTupleSlot != NULL)
191 :
192 : static TupleTableSlot *ExecHashJoinOuterGetTuple(PlanState *outerNode,
193 : HashJoinState *hjstate,
194 : uint32 *hashvalue);
195 : static TupleTableSlot *ExecParallelHashJoinOuterGetTuple(PlanState *outerNode,
196 : HashJoinState *hjstate,
197 : uint32 *hashvalue);
198 : static TupleTableSlot *ExecHashJoinGetSavedTuple(HashJoinState *hjstate,
199 : BufFile *file,
200 : uint32 *hashvalue,
201 : TupleTableSlot *tupleSlot);
202 : static bool ExecHashJoinNewBatch(HashJoinState *hjstate);
203 : static bool ExecParallelHashJoinNewBatch(HashJoinState *hjstate);
204 : static void ExecParallelHashJoinPartitionOuter(HashJoinState *hjstate);
205 :
206 :
207 : /* ----------------------------------------------------------------
208 : * ExecHashJoinImpl
209 : *
210 : * This function implements the Hybrid Hashjoin algorithm. It is marked
211 : * with an always-inline attribute so that ExecHashJoin() and
212 : * ExecParallelHashJoin() can inline it. Compilers that respect the
213 : * attribute should create versions specialized for parallel == true and
214 : * parallel == false with unnecessary branches removed.
215 : *
216 : * Note: the relation we build hash table on is the "inner"
217 : * the other one is "outer".
218 : * ----------------------------------------------------------------
219 : */
220 : static pg_attribute_always_inline TupleTableSlot *
221 9439630 : ExecHashJoinImpl(PlanState *pstate, bool parallel)
222 : {
223 9439630 : HashJoinState *node = castNode(HashJoinState, pstate);
224 : PlanState *outerNode;
225 : HashState *hashNode;
226 : ExprState *joinqual;
227 : ExprState *otherqual;
228 : ExprContext *econtext;
229 : HashJoinTable hashtable;
230 : TupleTableSlot *outerTupleSlot;
231 : uint32 hashvalue;
232 : int batchno;
233 : ParallelHashJoinState *parallel_state;
234 :
235 : /*
236 : * get information from HashJoin node
237 : */
238 9439630 : joinqual = node->js.joinqual;
239 9439630 : otherqual = node->js.ps.qual;
240 9439630 : hashNode = (HashState *) innerPlanState(node);
241 9439630 : outerNode = outerPlanState(node);
242 9439630 : hashtable = node->hj_HashTable;
243 9439630 : econtext = node->js.ps.ps_ExprContext;
244 9439630 : parallel_state = hashNode->parallel_state;
245 :
246 : /*
247 : * Reset per-tuple memory context to free any expression evaluation
248 : * storage allocated in the previous tuple cycle.
249 : */
250 9439630 : ResetExprContext(econtext);
251 :
252 : /*
253 : * run the hash join state machine
254 : */
255 : for (;;)
256 : {
257 : /*
258 : * It's possible to iterate this loop many times before returning a
259 : * tuple, in some pathological cases such as needing to move much of
260 : * the current batch to a later batch. So let's check for interrupts
261 : * each time through.
262 : */
263 36310416 : CHECK_FOR_INTERRUPTS();
264 :
265 36310416 : switch (node->hj_JoinState)
266 : {
267 24986 : case HJ_BUILD_HASHTABLE:
268 :
269 : /*
270 : * First time through: build hash table for inner relation.
271 : */
272 : Assert(hashtable == NULL);
273 :
274 : /*
275 : * If the outer relation is completely empty, and it's not
276 : * right/right-anti/full join, we can quit without building
277 : * the hash table. However, for an inner join it is only a
278 : * win to check this when the outer relation's startup cost is
279 : * less than the projected cost of building the hash table.
280 : * Otherwise it's best to build the hash table first and see
281 : * if the inner relation is empty. (When it's a left join, we
282 : * should always make this check, since we aren't going to be
283 : * able to skip the join on the strength of an empty inner
284 : * relation anyway.)
285 : *
286 : * If we are rescanning the join, we make use of information
287 : * gained on the previous scan: don't bother to try the
288 : * prefetch if the previous scan found the outer relation
289 : * nonempty. This is not 100% reliable since with new
290 : * parameters the outer relation might yield different
291 : * results, but it's a good heuristic.
292 : *
293 : * The only way to make the check is to try to fetch a tuple
294 : * from the outer plan node. If we succeed, we have to stash
295 : * it away for later consumption by ExecHashJoinOuterGetTuple.
296 : */
297 24986 : if (HJ_FILL_INNER(node))
298 : {
299 : /* no chance to not build the hash table */
300 6114 : node->hj_FirstOuterTupleSlot = NULL;
301 : }
302 18872 : else if (parallel)
303 : {
304 : /*
305 : * The empty-outer optimization is not implemented for
306 : * shared hash tables, because no one participant can
307 : * determine that there are no outer tuples, and it's not
308 : * yet clear that it's worth the synchronization overhead
309 : * of reaching consensus to figure that out. So we have
310 : * to build the hash table.
311 : */
312 326 : node->hj_FirstOuterTupleSlot = NULL;
313 : }
314 18546 : else if (HJ_FILL_OUTER(node) ||
315 13880 : (outerNode->plan->startup_cost < hashNode->ps.plan->total_cost &&
316 12890 : !node->hj_OuterNotEmpty))
317 : {
318 16604 : node->hj_FirstOuterTupleSlot = ExecProcNode(outerNode);
319 16604 : if (TupIsNull(node->hj_FirstOuterTupleSlot))
320 : {
321 4036 : node->hj_OuterNotEmpty = false;
322 4036 : return NULL;
323 : }
324 : else
325 12568 : node->hj_OuterNotEmpty = true;
326 : }
327 : else
328 1942 : node->hj_FirstOuterTupleSlot = NULL;
329 :
330 : /*
331 : * Create the hash table. If using Parallel Hash, then
332 : * whoever gets here first will create the hash table and any
333 : * later arrivals will merely attach to it.
334 : */
335 20950 : hashtable = ExecHashTableCreate(hashNode);
336 20950 : node->hj_HashTable = hashtable;
337 :
338 : /*
339 : * Execute the Hash node, to build the hash table. If using
340 : * Parallel Hash, then we'll try to help hashing unless we
341 : * arrived too late.
342 : */
343 20950 : hashNode->hashtable = hashtable;
344 20950 : (void) MultiExecProcNode((PlanState *) hashNode);
345 :
346 : /*
347 : * If the inner relation is completely empty, and we're not
348 : * doing a left outer join, we can quit without scanning the
349 : * outer relation.
350 : */
351 20950 : if (hashtable->totalTuples == 0 && !HJ_FILL_OUTER(node))
352 : {
353 1222 : if (parallel)
354 : {
355 : /*
356 : * Advance the build barrier to PHJ_BUILD_RUN before
357 : * proceeding so we can negotiate resource cleanup.
358 : */
359 6 : Barrier *build_barrier = ¶llel_state->build_barrier;
360 :
361 8 : while (BarrierPhase(build_barrier) < PHJ_BUILD_RUN)
362 2 : BarrierArriveAndWait(build_barrier, 0);
363 : }
364 1222 : return NULL;
365 : }
366 :
367 : /*
368 : * need to remember whether nbatch has increased since we
369 : * began scanning the outer relation
370 : */
371 19728 : hashtable->nbatch_outstart = hashtable->nbatch;
372 :
373 : /*
374 : * Reset OuterNotEmpty for scan. (It's OK if we fetched a
375 : * tuple above, because ExecHashJoinOuterGetTuple will
376 : * immediately set it again.)
377 : */
378 19728 : node->hj_OuterNotEmpty = false;
379 :
380 19728 : if (parallel)
381 : {
382 : Barrier *build_barrier;
383 :
384 392 : build_barrier = ¶llel_state->build_barrier;
385 : Assert(BarrierPhase(build_barrier) == PHJ_BUILD_HASH_OUTER ||
386 : BarrierPhase(build_barrier) == PHJ_BUILD_RUN ||
387 : BarrierPhase(build_barrier) == PHJ_BUILD_FREE);
388 392 : if (BarrierPhase(build_barrier) == PHJ_BUILD_HASH_OUTER)
389 : {
390 : /*
391 : * If multi-batch, we need to hash the outer relation
392 : * up front.
393 : */
394 262 : if (hashtable->nbatch > 1)
395 146 : ExecParallelHashJoinPartitionOuter(node);
396 262 : BarrierArriveAndWait(build_barrier,
397 : WAIT_EVENT_HASH_BUILD_HASH_OUTER);
398 : }
399 130 : else if (BarrierPhase(build_barrier) == PHJ_BUILD_FREE)
400 : {
401 : /*
402 : * If we attached so late that the job is finished and
403 : * the batch state has been freed, we can return
404 : * immediately.
405 : */
406 0 : return NULL;
407 : }
408 :
409 : /* Each backend should now select a batch to work on. */
410 : Assert(BarrierPhase(build_barrier) == PHJ_BUILD_RUN);
411 392 : hashtable->curbatch = -1;
412 392 : node->hj_JoinState = HJ_NEED_NEW_BATCH;
413 :
414 392 : continue;
415 : }
416 : else
417 19336 : node->hj_JoinState = HJ_NEED_NEW_OUTER;
418 :
419 : /* FALL THRU */
420 :
421 17144342 : case HJ_NEED_NEW_OUTER:
422 :
423 : /*
424 : * We don't have an outer tuple, try to get the next one
425 : */
426 17144342 : if (parallel)
427 : outerTupleSlot =
428 2160896 : ExecParallelHashJoinOuterGetTuple(outerNode, node,
429 : &hashvalue);
430 : else
431 : outerTupleSlot =
432 14983446 : ExecHashJoinOuterGetTuple(outerNode, node, &hashvalue);
433 :
434 17144342 : if (TupIsNull(outerTupleSlot))
435 : {
436 : /* end of batch, or maybe whole join */
437 22002 : if (HJ_FILL_INNER(node))
438 : {
439 : /* set up to scan for unmatched inner tuples */
440 5838 : if (parallel)
441 : {
442 : /*
443 : * Only one process is currently allow to handle
444 : * each batch's unmatched tuples, in a parallel
445 : * join.
446 : */
447 72 : if (ExecParallelPrepHashTableForUnmatched(node))
448 66 : node->hj_JoinState = HJ_FILL_INNER_TUPLES;
449 : else
450 6 : node->hj_JoinState = HJ_NEED_NEW_BATCH;
451 : }
452 : else
453 : {
454 5766 : ExecPrepHashTableForUnmatched(node);
455 5766 : node->hj_JoinState = HJ_FILL_INNER_TUPLES;
456 : }
457 : }
458 : else
459 16164 : node->hj_JoinState = HJ_NEED_NEW_BATCH;
460 22002 : continue;
461 : }
462 :
463 17122340 : econtext->ecxt_outertuple = outerTupleSlot;
464 17122340 : node->hj_MatchedOuter = false;
465 :
466 : /*
467 : * Find the corresponding bucket for this tuple in the main
468 : * hash table or skew hash table.
469 : */
470 17122340 : node->hj_CurHashValue = hashvalue;
471 17122340 : ExecHashGetBucketAndBatch(hashtable, hashvalue,
472 : &node->hj_CurBucketNo, &batchno);
473 17122340 : node->hj_CurSkewBucketNo = ExecHashGetSkewBucket(hashtable,
474 : hashvalue);
475 17122340 : node->hj_CurTuple = NULL;
476 :
477 : /*
478 : * The tuple might not belong to the current batch (where
479 : * "current batch" includes the skew buckets if any).
480 : */
481 17122340 : if (batchno != hashtable->curbatch &&
482 1471392 : node->hj_CurSkewBucketNo == INVALID_SKEW_BUCKET_NO)
483 : {
484 : bool shouldFree;
485 1470192 : MinimalTuple mintuple = ExecFetchSlotMinimalTuple(outerTupleSlot,
486 : &shouldFree);
487 :
488 : /*
489 : * Need to postpone this outer tuple to a later batch.
490 : * Save it in the corresponding outer-batch file.
491 : */
492 : Assert(parallel_state == NULL);
493 : Assert(batchno > hashtable->curbatch);
494 1470192 : ExecHashJoinSaveTuple(mintuple, hashvalue,
495 1470192 : &hashtable->outerBatchFile[batchno],
496 : hashtable);
497 :
498 1470192 : if (shouldFree)
499 1470192 : heap_free_minimal_tuple(mintuple);
500 :
501 : /* Loop around, staying in HJ_NEED_NEW_OUTER state */
502 1470192 : continue;
503 : }
504 :
505 : /* OK, let's scan the bucket for matches */
506 15652148 : node->hj_JoinState = HJ_SCAN_BUCKET;
507 :
508 : /* FALL THRU */
509 :
510 22113472 : case HJ_SCAN_BUCKET:
511 :
512 : /*
513 : * Scan the selected hash bucket for matches to current outer
514 : */
515 22113472 : if (parallel)
516 : {
517 4200054 : if (!ExecParallelScanHashBucket(node, econtext))
518 : {
519 : /* out of matches; check for possible outer-join fill */
520 2160030 : node->hj_JoinState = HJ_FILL_OUTER_TUPLE;
521 2160030 : continue;
522 : }
523 : }
524 : else
525 : {
526 17913418 : if (!ExecScanHashBucket(node, econtext))
527 : {
528 : /* out of matches; check for possible outer-join fill */
529 10062500 : node->hj_JoinState = HJ_FILL_OUTER_TUPLE;
530 10062500 : continue;
531 : }
532 : }
533 :
534 : /*
535 : * In a right-semijoin, we only need the first match for each
536 : * inner tuple.
537 : */
538 9890942 : if (node->js.jointype == JOIN_RIGHT_SEMI &&
539 304 : HeapTupleHeaderHasMatch(HJTUPLE_MINTUPLE(node->hj_CurTuple)))
540 72 : continue;
541 :
542 : /*
543 : * We've got a match, but still need to test non-hashed quals.
544 : * ExecScanHashBucket already set up all the state needed to
545 : * call ExecQual.
546 : *
547 : * If we pass the qual, then save state for next call and have
548 : * ExecProject form the projection, store it in the tuple
549 : * table, and return the slot.
550 : *
551 : * Only the joinquals determine tuple match status, but all
552 : * quals must pass to actually return the tuple.
553 : */
554 9890870 : if (joinqual == NULL || ExecQual(joinqual, econtext))
555 : {
556 9738440 : node->hj_MatchedOuter = true;
557 :
558 : /*
559 : * This is really only needed if HJ_FILL_INNER(node) or if
560 : * we are in a right-semijoin, but we'll avoid the branch
561 : * and just set it always.
562 : */
563 9738440 : if (!HeapTupleHeaderHasMatch(HJTUPLE_MINTUPLE(node->hj_CurTuple)))
564 5793036 : HeapTupleHeaderSetMatch(HJTUPLE_MINTUPLE(node->hj_CurTuple));
565 :
566 : /* In an antijoin, we never return a matched tuple */
567 9738440 : if (node->js.jointype == JOIN_ANTI)
568 : {
569 1556360 : node->hj_JoinState = HJ_NEED_NEW_OUTER;
570 1556360 : continue;
571 : }
572 :
573 : /*
574 : * If we only need to consider the first matching inner
575 : * tuple, then advance to next outer tuple after we've
576 : * processed this one.
577 : */
578 8182080 : if (node->js.single_match)
579 1873132 : node->hj_JoinState = HJ_NEED_NEW_OUTER;
580 :
581 : /*
582 : * In a right-antijoin, we never return a matched tuple.
583 : * If it's not an inner_unique join, we need to stay on
584 : * the current outer tuple to continue scanning the inner
585 : * side for matches.
586 : */
587 8182080 : if (node->js.jointype == JOIN_RIGHT_ANTI)
588 26116 : continue;
589 :
590 8155964 : if (otherqual == NULL || ExecQual(otherqual, econtext))
591 7971154 : return ExecProject(node->js.ps.ps_ProjInfo);
592 : else
593 184810 : InstrCountFiltered2(node, 1);
594 : }
595 : else
596 152430 : InstrCountFiltered1(node, 1);
597 337240 : break;
598 :
599 12222530 : case HJ_FILL_OUTER_TUPLE:
600 :
601 : /*
602 : * The current outer tuple has run out of matches, so check
603 : * whether to emit a dummy outer-join tuple. Whether we emit
604 : * one or not, the next state is NEED_NEW_OUTER.
605 : */
606 12222530 : node->hj_JoinState = HJ_NEED_NEW_OUTER;
607 :
608 12222530 : if (!node->hj_MatchedOuter &&
609 7246886 : HJ_FILL_OUTER(node))
610 : {
611 : /*
612 : * Generate a fake join tuple with nulls for the inner
613 : * tuple, and return it if it passes the non-join quals.
614 : */
615 2088968 : econtext->ecxt_innertuple = node->hj_NullInnerTupleSlot;
616 :
617 2088968 : if (otherqual == NULL || ExecQual(otherqual, econtext))
618 1001732 : return ExecProject(node->js.ps.ps_ProjInfo);
619 : else
620 1087236 : InstrCountFiltered2(node, 1);
621 : }
622 11220798 : break;
623 :
624 454188 : case HJ_FILL_INNER_TUPLES:
625 :
626 : /*
627 : * We have finished a batch, but we are doing
628 : * right/right-anti/full join, so any unmatched inner tuples
629 : * in the hashtable have to be emitted before we continue to
630 : * the next batch.
631 : */
632 788304 : if (!(parallel ? ExecParallelScanHashTableForUnmatched(node, econtext)
633 334116 : : ExecScanHashTableForUnmatched(node, econtext)))
634 : {
635 : /* no more unmatched tuples */
636 5820 : node->hj_JoinState = HJ_NEED_NEW_BATCH;
637 5820 : continue;
638 : }
639 :
640 : /*
641 : * Generate a fake join tuple with nulls for the outer tuple,
642 : * and return it if it passes the non-join quals.
643 : */
644 448368 : econtext->ecxt_outertuple = node->hj_NullOuterTupleSlot;
645 :
646 448368 : if (otherqual == NULL || ExecQual(otherqual, econtext))
647 441210 : return ExecProject(node->js.ps.ps_ProjInfo);
648 : else
649 7158 : InstrCountFiltered2(node, 1);
650 7158 : break;
651 :
652 22382 : case HJ_NEED_NEW_BATCH:
653 :
654 : /*
655 : * Try to advance to next batch. Done if there are no more.
656 : */
657 22382 : if (parallel)
658 : {
659 1258 : if (!ExecParallelHashJoinNewBatch(node))
660 392 : return NULL; /* end of parallel-aware join */
661 : }
662 : else
663 : {
664 21124 : if (!ExecHashJoinNewBatch(node))
665 19884 : return NULL; /* end of parallel-oblivious join */
666 : }
667 2106 : node->hj_JoinState = HJ_NEED_NEW_OUTER;
668 2106 : break;
669 :
670 0 : default:
671 0 : elog(ERROR, "unrecognized hashjoin state: %d",
672 : (int) node->hj_JoinState);
673 : }
674 : }
675 : }
676 :
677 : /* ----------------------------------------------------------------
678 : * ExecHashJoin
679 : *
680 : * Parallel-oblivious version.
681 : * ----------------------------------------------------------------
682 : */
683 : static TupleTableSlot * /* return: a tuple or NULL */
684 7159196 : ExecHashJoin(PlanState *pstate)
685 : {
686 : /*
687 : * On sufficiently smart compilers this should be inlined with the
688 : * parallel-aware branches removed.
689 : */
690 7159196 : return ExecHashJoinImpl(pstate, false);
691 : }
692 :
693 : /* ----------------------------------------------------------------
694 : * ExecParallelHashJoin
695 : *
696 : * Parallel-aware version.
697 : * ----------------------------------------------------------------
698 : */
699 : static TupleTableSlot * /* return: a tuple or NULL */
700 2280434 : ExecParallelHashJoin(PlanState *pstate)
701 : {
702 : /*
703 : * On sufficiently smart compilers this should be inlined with the
704 : * parallel-oblivious branches removed.
705 : */
706 2280434 : return ExecHashJoinImpl(pstate, true);
707 : }
708 :
709 : /* ----------------------------------------------------------------
710 : * ExecInitHashJoin
711 : *
712 : * Init routine for HashJoin node.
713 : * ----------------------------------------------------------------
714 : */
715 : HashJoinState *
716 31270 : ExecInitHashJoin(HashJoin *node, EState *estate, int eflags)
717 : {
718 : HashJoinState *hjstate;
719 : Plan *outerNode;
720 : Hash *hashNode;
721 : TupleDesc outerDesc,
722 : innerDesc;
723 : const TupleTableSlotOps *ops;
724 :
725 : /* check for unsupported flags */
726 : Assert(!(eflags & (EXEC_FLAG_BACKWARD | EXEC_FLAG_MARK)));
727 :
728 : /*
729 : * create state structure
730 : */
731 31270 : hjstate = makeNode(HashJoinState);
732 31270 : hjstate->js.ps.plan = (Plan *) node;
733 31270 : hjstate->js.ps.state = estate;
734 :
735 : /*
736 : * See ExecHashJoinInitializeDSM() and ExecHashJoinInitializeWorker()
737 : * where this function may be replaced with a parallel version, if we
738 : * managed to launch a parallel query.
739 : */
740 31270 : hjstate->js.ps.ExecProcNode = ExecHashJoin;
741 31270 : hjstate->js.jointype = node->join.jointype;
742 :
743 : /*
744 : * Miscellaneous initialization
745 : *
746 : * create expression context for node
747 : */
748 31270 : ExecAssignExprContext(estate, &hjstate->js.ps);
749 :
750 : /*
751 : * initialize child nodes
752 : *
753 : * Note: we could suppress the REWIND flag for the inner input, which
754 : * would amount to betting that the hash will be a single batch. Not
755 : * clear if this would be a win or not.
756 : */
757 31270 : outerNode = outerPlan(node);
758 31270 : hashNode = (Hash *) innerPlan(node);
759 :
760 31270 : outerPlanState(hjstate) = ExecInitNode(outerNode, estate, eflags);
761 31270 : outerDesc = ExecGetResultType(outerPlanState(hjstate));
762 31270 : innerPlanState(hjstate) = ExecInitNode((Plan *) hashNode, estate, eflags);
763 31270 : innerDesc = ExecGetResultType(innerPlanState(hjstate));
764 :
765 : /*
766 : * Initialize result slot, type and projection.
767 : */
768 31270 : ExecInitResultTupleSlotTL(&hjstate->js.ps, &TTSOpsVirtual);
769 31270 : ExecAssignProjectionInfo(&hjstate->js.ps, NULL);
770 :
771 : /*
772 : * tuple table initialization
773 : */
774 31270 : ops = ExecGetResultSlotOps(outerPlanState(hjstate), NULL);
775 31270 : hjstate->hj_OuterTupleSlot = ExecInitExtraTupleSlot(estate, outerDesc,
776 : ops);
777 :
778 : /*
779 : * detect whether we need only consider the first matching inner tuple
780 : */
781 46478 : hjstate->js.single_match = (node->join.inner_unique ||
782 15208 : node->join.jointype == JOIN_SEMI);
783 :
784 : /* set up null tuples for outer joins, if needed */
785 31270 : switch (node->join.jointype)
786 : {
787 18928 : case JOIN_INNER:
788 : case JOIN_SEMI:
789 : case JOIN_RIGHT_SEMI:
790 18928 : break;
791 5116 : case JOIN_LEFT:
792 : case JOIN_ANTI:
793 5116 : hjstate->hj_NullInnerTupleSlot =
794 5116 : ExecInitNullTupleSlot(estate, innerDesc, &TTSOpsVirtual);
795 5116 : break;
796 6130 : case JOIN_RIGHT:
797 : case JOIN_RIGHT_ANTI:
798 6130 : hjstate->hj_NullOuterTupleSlot =
799 6130 : ExecInitNullTupleSlot(estate, outerDesc, &TTSOpsVirtual);
800 6130 : break;
801 1096 : case JOIN_FULL:
802 1096 : hjstate->hj_NullOuterTupleSlot =
803 1096 : ExecInitNullTupleSlot(estate, outerDesc, &TTSOpsVirtual);
804 1096 : hjstate->hj_NullInnerTupleSlot =
805 1096 : ExecInitNullTupleSlot(estate, innerDesc, &TTSOpsVirtual);
806 1096 : break;
807 0 : default:
808 0 : elog(ERROR, "unrecognized join type: %d",
809 : (int) node->join.jointype);
810 : }
811 :
812 : /*
813 : * now for some voodoo. our temporary tuple slot is actually the result
814 : * tuple slot of the Hash node (which is our inner plan). we can do this
815 : * because Hash nodes don't return tuples via ExecProcNode() -- instead
816 : * the hash join node uses ExecScanHashBucket() to get at the contents of
817 : * the hash table. -cim 6/9/91
818 : */
819 : {
820 31270 : HashState *hashstate = (HashState *) innerPlanState(hjstate);
821 31270 : Hash *hash = (Hash *) hashstate->ps.plan;
822 31270 : TupleTableSlot *slot = hashstate->ps.ps_ResultTupleSlot;
823 : Oid *outer_hashfuncid;
824 : Oid *inner_hashfuncid;
825 : bool *hash_strict;
826 : ListCell *lc;
827 : int nkeys;
828 :
829 :
830 31270 : hjstate->hj_HashTupleSlot = slot;
831 :
832 : /*
833 : * Build ExprStates to obtain hash values for either side of the join.
834 : * This must be done here as ExecBuildHash32Expr needs to know how to
835 : * handle NULL inputs and the required handling of that depends on the
836 : * jointype. We don't know the join type in ExecInitHash() and we
837 : * must build the ExprStates before ExecHashTableCreate() so we
838 : * properly attribute any SubPlans that exist in the hash expressions
839 : * to the correct PlanState.
840 : */
841 31270 : nkeys = list_length(node->hashoperators);
842 :
843 31270 : outer_hashfuncid = palloc_array(Oid, nkeys);
844 31270 : inner_hashfuncid = palloc_array(Oid, nkeys);
845 31270 : hash_strict = palloc_array(bool, nkeys);
846 :
847 : /*
848 : * Determine the hash function for each side of the join for the given
849 : * hash operator.
850 : */
851 64960 : foreach(lc, node->hashoperators)
852 : {
853 33690 : Oid hashop = lfirst_oid(lc);
854 33690 : int i = foreach_current_index(lc);
855 :
856 33690 : if (!get_op_hash_functions(hashop,
857 33690 : &outer_hashfuncid[i],
858 33690 : &inner_hashfuncid[i]))
859 0 : elog(ERROR,
860 : "could not find hash function for hash operator %u",
861 : hashop);
862 33690 : hash_strict[i] = op_strict(hashop);
863 : }
864 :
865 : /*
866 : * Build an ExprState to generate the hash value for the expressions
867 : * on the outer of the join. This ExprState must finish generating
868 : * the hash value when HJ_FILL_OUTER() is true. Otherwise,
869 : * ExecBuildHash32Expr will set up the ExprState to abort early if it
870 : * finds a NULL. In these cases, we don't need to store these tuples
871 : * in the hash table as the jointype does not require it.
872 : */
873 31270 : hjstate->hj_OuterHash =
874 31270 : ExecBuildHash32Expr(hjstate->js.ps.ps_ResultTupleDesc,
875 : hjstate->js.ps.resultops,
876 : outer_hashfuncid,
877 31270 : node->hashcollations,
878 31270 : node->hashkeys,
879 : hash_strict,
880 : &hjstate->js.ps,
881 : 0,
882 31270 : HJ_FILL_OUTER(hjstate));
883 :
884 : /* As above, but for the inner side of the join */
885 31270 : hashstate->hash_expr =
886 31270 : ExecBuildHash32Expr(hashstate->ps.ps_ResultTupleDesc,
887 : hashstate->ps.resultops,
888 : inner_hashfuncid,
889 31270 : node->hashcollations,
890 31270 : hash->hashkeys,
891 : hash_strict,
892 : &hashstate->ps,
893 : 0,
894 31270 : HJ_FILL_INNER(hjstate));
895 :
896 : /*
897 : * Set up the skew table hash function while we have a record of the
898 : * first key's hash function Oid.
899 : */
900 31270 : if (OidIsValid(hash->skewTable))
901 : {
902 22702 : hashstate->skew_hashfunction = palloc0(sizeof(FmgrInfo));
903 22702 : hashstate->skew_collation = linitial_oid(node->hashcollations);
904 22702 : fmgr_info(outer_hashfuncid[0], hashstate->skew_hashfunction);
905 : }
906 :
907 : /* no need to keep these */
908 31270 : pfree(outer_hashfuncid);
909 31270 : pfree(inner_hashfuncid);
910 31270 : pfree(hash_strict);
911 : }
912 :
913 : /*
914 : * initialize child expressions
915 : */
916 31270 : hjstate->js.ps.qual =
917 31270 : ExecInitQual(node->join.plan.qual, (PlanState *) hjstate);
918 31270 : hjstate->js.joinqual =
919 31270 : ExecInitQual(node->join.joinqual, (PlanState *) hjstate);
920 31270 : hjstate->hashclauses =
921 31270 : ExecInitQual(node->hashclauses, (PlanState *) hjstate);
922 :
923 : /*
924 : * initialize hash-specific info
925 : */
926 31270 : hjstate->hj_HashTable = NULL;
927 31270 : hjstate->hj_FirstOuterTupleSlot = NULL;
928 :
929 31270 : hjstate->hj_CurHashValue = 0;
930 31270 : hjstate->hj_CurBucketNo = 0;
931 31270 : hjstate->hj_CurSkewBucketNo = INVALID_SKEW_BUCKET_NO;
932 31270 : hjstate->hj_CurTuple = NULL;
933 :
934 31270 : hjstate->hj_JoinState = HJ_BUILD_HASHTABLE;
935 31270 : hjstate->hj_MatchedOuter = false;
936 31270 : hjstate->hj_OuterNotEmpty = false;
937 :
938 31270 : return hjstate;
939 : }
940 :
941 : /* ----------------------------------------------------------------
942 : * ExecEndHashJoin
943 : *
944 : * clean up routine for HashJoin node
945 : * ----------------------------------------------------------------
946 : */
947 : void
948 31164 : ExecEndHashJoin(HashJoinState *node)
949 : {
950 : /*
951 : * Free hash table
952 : */
953 31164 : if (node->hj_HashTable)
954 : {
955 19756 : ExecHashTableDestroy(node->hj_HashTable);
956 19756 : node->hj_HashTable = NULL;
957 : }
958 :
959 : /*
960 : * clean up subtrees
961 : */
962 31164 : ExecEndNode(outerPlanState(node));
963 31164 : ExecEndNode(innerPlanState(node));
964 31164 : }
965 :
966 : /*
967 : * ExecHashJoinOuterGetTuple
968 : *
969 : * get the next outer tuple for a parallel oblivious hashjoin: either by
970 : * executing the outer plan node in the first pass, or from the temp
971 : * files for the hashjoin batches.
972 : *
973 : * Returns a null slot if no more outer tuples (within the current batch).
974 : *
975 : * On success, the tuple's hash value is stored at *hashvalue --- this is
976 : * either originally computed, or re-read from the temp file.
977 : */
978 : static TupleTableSlot *
979 14983446 : ExecHashJoinOuterGetTuple(PlanState *outerNode,
980 : HashJoinState *hjstate,
981 : uint32 *hashvalue)
982 : {
983 14983446 : HashJoinTable hashtable = hjstate->hj_HashTable;
984 14983446 : int curbatch = hashtable->curbatch;
985 : TupleTableSlot *slot;
986 :
987 14983446 : if (curbatch == 0) /* if it is the first pass */
988 : {
989 : /*
990 : * Check to see if first outer tuple was already fetched by
991 : * ExecHashJoin() and not used yet.
992 : */
993 13512014 : slot = hjstate->hj_FirstOuterTupleSlot;
994 13512014 : if (!TupIsNull(slot))
995 11890 : hjstate->hj_FirstOuterTupleSlot = NULL;
996 : else
997 13500124 : slot = ExecProcNode(outerNode);
998 :
999 13512822 : while (!TupIsNull(slot))
1000 : {
1001 : bool isnull;
1002 :
1003 : /*
1004 : * We have to compute the tuple's hash value.
1005 : */
1006 13492926 : ExprContext *econtext = hjstate->js.ps.ps_ExprContext;
1007 :
1008 13492926 : econtext->ecxt_outertuple = slot;
1009 :
1010 13492926 : ResetExprContext(econtext);
1011 :
1012 13492926 : *hashvalue = DatumGetUInt32(ExecEvalExprSwitchContext(hjstate->hj_OuterHash,
1013 : econtext,
1014 : &isnull));
1015 :
1016 13492926 : if (!isnull)
1017 : {
1018 : /* remember outer relation is not empty for possible rescan */
1019 13492118 : hjstate->hj_OuterNotEmpty = true;
1020 :
1021 13492118 : return slot;
1022 : }
1023 :
1024 : /*
1025 : * That tuple couldn't match because of a NULL, so discard it and
1026 : * continue with the next one.
1027 : */
1028 808 : slot = ExecProcNode(outerNode);
1029 : }
1030 : }
1031 1471432 : else if (curbatch < hashtable->nbatch)
1032 : {
1033 1471432 : BufFile *file = hashtable->outerBatchFile[curbatch];
1034 :
1035 : /*
1036 : * In outer-join cases, we could get here even though the batch file
1037 : * is empty.
1038 : */
1039 1471432 : if (file == NULL)
1040 0 : return NULL;
1041 :
1042 1471432 : slot = ExecHashJoinGetSavedTuple(hjstate,
1043 : file,
1044 : hashvalue,
1045 : hjstate->hj_OuterTupleSlot);
1046 1471432 : if (!TupIsNull(slot))
1047 1470192 : return slot;
1048 : }
1049 :
1050 : /* End of this batch */
1051 21136 : return NULL;
1052 : }
1053 :
1054 : /*
1055 : * ExecHashJoinOuterGetTuple variant for the parallel case.
1056 : */
1057 : static TupleTableSlot *
1058 2160896 : ExecParallelHashJoinOuterGetTuple(PlanState *outerNode,
1059 : HashJoinState *hjstate,
1060 : uint32 *hashvalue)
1061 : {
1062 2160896 : HashJoinTable hashtable = hjstate->hj_HashTable;
1063 2160896 : int curbatch = hashtable->curbatch;
1064 : TupleTableSlot *slot;
1065 :
1066 : /*
1067 : * In the Parallel Hash case we only run the outer plan directly for
1068 : * single-batch hash joins. Otherwise we have to go to batch files, even
1069 : * for batch 0.
1070 : */
1071 2160896 : if (curbatch == 0 && hashtable->nbatch == 1)
1072 : {
1073 960124 : slot = ExecProcNode(outerNode);
1074 :
1075 960124 : while (!TupIsNull(slot))
1076 : {
1077 : bool isnull;
1078 :
1079 960006 : ExprContext *econtext = hjstate->js.ps.ps_ExprContext;
1080 :
1081 960006 : econtext->ecxt_outertuple = slot;
1082 :
1083 960006 : ResetExprContext(econtext);
1084 :
1085 960006 : *hashvalue = DatumGetUInt32(ExecEvalExprSwitchContext(hjstate->hj_OuterHash,
1086 : econtext,
1087 : &isnull));
1088 :
1089 960006 : if (!isnull)
1090 960006 : return slot;
1091 :
1092 : /*
1093 : * That tuple couldn't match because of a NULL, so discard it and
1094 : * continue with the next one.
1095 : */
1096 0 : slot = ExecProcNode(outerNode);
1097 : }
1098 : }
1099 1200772 : else if (curbatch < hashtable->nbatch)
1100 : {
1101 : MinimalTuple tuple;
1102 :
1103 1200772 : tuple = sts_parallel_scan_next(hashtable->batches[curbatch].outer_tuples,
1104 : hashvalue);
1105 1200772 : if (tuple != NULL)
1106 : {
1107 1200024 : ExecForceStoreMinimalTuple(tuple,
1108 : hjstate->hj_OuterTupleSlot,
1109 : false);
1110 1200024 : slot = hjstate->hj_OuterTupleSlot;
1111 1200024 : return slot;
1112 : }
1113 : else
1114 748 : ExecClearTuple(hjstate->hj_OuterTupleSlot);
1115 : }
1116 :
1117 : /* End of this batch */
1118 866 : hashtable->batches[curbatch].outer_eof = true;
1119 :
1120 866 : return NULL;
1121 : }
1122 :
1123 : /*
1124 : * ExecHashJoinNewBatch
1125 : * switch to a new hashjoin batch
1126 : *
1127 : * Returns true if successful, false if there are no more batches.
1128 : */
1129 : static bool
1130 21124 : ExecHashJoinNewBatch(HashJoinState *hjstate)
1131 : {
1132 21124 : HashJoinTable hashtable = hjstate->hj_HashTable;
1133 : int nbatch;
1134 : int curbatch;
1135 : BufFile *innerFile;
1136 : TupleTableSlot *slot;
1137 : uint32 hashvalue;
1138 :
1139 21124 : nbatch = hashtable->nbatch;
1140 21124 : curbatch = hashtable->curbatch;
1141 :
1142 21124 : if (curbatch > 0)
1143 : {
1144 : /*
1145 : * We no longer need the previous outer batch file; close it right
1146 : * away to free disk space.
1147 : */
1148 1240 : if (hashtable->outerBatchFile[curbatch])
1149 1240 : BufFileClose(hashtable->outerBatchFile[curbatch]);
1150 1240 : hashtable->outerBatchFile[curbatch] = NULL;
1151 : }
1152 : else /* we just finished the first batch */
1153 : {
1154 : /*
1155 : * Reset some of the skew optimization state variables, since we no
1156 : * longer need to consider skew tuples after the first batch. The
1157 : * memory context reset we are about to do will release the skew
1158 : * hashtable itself.
1159 : */
1160 19884 : hashtable->skewEnabled = false;
1161 19884 : hashtable->skewBucket = NULL;
1162 19884 : hashtable->skewBucketNums = NULL;
1163 19884 : hashtable->nSkewBuckets = 0;
1164 19884 : hashtable->spaceUsedSkew = 0;
1165 : }
1166 :
1167 : /*
1168 : * We can always skip over any batches that are completely empty on both
1169 : * sides. We can sometimes skip over batches that are empty on only one
1170 : * side, but there are exceptions:
1171 : *
1172 : * 1. In a left/full outer join, we have to process outer batches even if
1173 : * the inner batch is empty. Similarly, in a right/right-anti/full outer
1174 : * join, we have to process inner batches even if the outer batch is
1175 : * empty.
1176 : *
1177 : * 2. If we have increased nbatch since the initial estimate, we have to
1178 : * scan inner batches since they might contain tuples that need to be
1179 : * reassigned to later inner batches.
1180 : *
1181 : * 3. Similarly, if we have increased nbatch since starting the outer
1182 : * scan, we have to rescan outer batches in case they contain tuples that
1183 : * need to be reassigned.
1184 : */
1185 21124 : curbatch++;
1186 21124 : while (curbatch < nbatch &&
1187 1240 : (hashtable->outerBatchFile[curbatch] == NULL ||
1188 1240 : hashtable->innerBatchFile[curbatch] == NULL))
1189 : {
1190 0 : if (hashtable->outerBatchFile[curbatch] &&
1191 0 : HJ_FILL_OUTER(hjstate))
1192 0 : break; /* must process due to rule 1 */
1193 0 : if (hashtable->innerBatchFile[curbatch] &&
1194 0 : HJ_FILL_INNER(hjstate))
1195 0 : break; /* must process due to rule 1 */
1196 0 : if (hashtable->innerBatchFile[curbatch] &&
1197 0 : nbatch != hashtable->nbatch_original)
1198 0 : break; /* must process due to rule 2 */
1199 0 : if (hashtable->outerBatchFile[curbatch] &&
1200 0 : nbatch != hashtable->nbatch_outstart)
1201 0 : break; /* must process due to rule 3 */
1202 : /* We can ignore this batch. */
1203 : /* Release associated temp files right away. */
1204 0 : if (hashtable->innerBatchFile[curbatch])
1205 0 : BufFileClose(hashtable->innerBatchFile[curbatch]);
1206 0 : hashtable->innerBatchFile[curbatch] = NULL;
1207 0 : if (hashtable->outerBatchFile[curbatch])
1208 0 : BufFileClose(hashtable->outerBatchFile[curbatch]);
1209 0 : hashtable->outerBatchFile[curbatch] = NULL;
1210 0 : curbatch++;
1211 : }
1212 :
1213 21124 : if (curbatch >= nbatch)
1214 19884 : return false; /* no more batches */
1215 :
1216 1240 : hashtable->curbatch = curbatch;
1217 :
1218 : /*
1219 : * Reload the hash table with the new inner batch (which could be empty)
1220 : */
1221 1240 : ExecHashTableReset(hashtable);
1222 :
1223 1240 : innerFile = hashtable->innerBatchFile[curbatch];
1224 :
1225 1240 : if (innerFile != NULL)
1226 : {
1227 1240 : if (BufFileSeek(innerFile, 0, 0, SEEK_SET))
1228 0 : ereport(ERROR,
1229 : (errcode_for_file_access(),
1230 : errmsg("could not rewind hash-join temporary file")));
1231 :
1232 2545322 : while ((slot = ExecHashJoinGetSavedTuple(hjstate,
1233 : innerFile,
1234 : &hashvalue,
1235 : hjstate->hj_HashTupleSlot)))
1236 : {
1237 : /*
1238 : * NOTE: some tuples may be sent to future batches. Also, it is
1239 : * possible for hashtable->nbatch to be increased here!
1240 : */
1241 2544082 : ExecHashTableInsert(hashtable, slot, hashvalue);
1242 : }
1243 :
1244 : /*
1245 : * after we build the hash table, the inner batch file is no longer
1246 : * needed
1247 : */
1248 1240 : BufFileClose(innerFile);
1249 1240 : hashtable->innerBatchFile[curbatch] = NULL;
1250 : }
1251 :
1252 : /*
1253 : * Rewind outer batch file (if present), so that we can start reading it.
1254 : */
1255 1240 : if (hashtable->outerBatchFile[curbatch] != NULL)
1256 : {
1257 1240 : if (BufFileSeek(hashtable->outerBatchFile[curbatch], 0, 0, SEEK_SET))
1258 0 : ereport(ERROR,
1259 : (errcode_for_file_access(),
1260 : errmsg("could not rewind hash-join temporary file")));
1261 : }
1262 :
1263 1240 : return true;
1264 : }
1265 :
1266 : /*
1267 : * Choose a batch to work on, and attach to it. Returns true if successful,
1268 : * false if there are no more batches.
1269 : */
1270 : static bool
1271 1258 : ExecParallelHashJoinNewBatch(HashJoinState *hjstate)
1272 : {
1273 1258 : HashJoinTable hashtable = hjstate->hj_HashTable;
1274 : int start_batchno;
1275 : int batchno;
1276 :
1277 : /*
1278 : * If we were already attached to a batch, remember not to bother checking
1279 : * it again, and detach from it (possibly freeing the hash table if we are
1280 : * last to detach).
1281 : */
1282 1258 : if (hashtable->curbatch >= 0)
1283 : {
1284 860 : hashtable->batches[hashtable->curbatch].done = true;
1285 860 : ExecHashTableDetachBatch(hashtable);
1286 : }
1287 :
1288 : /*
1289 : * Search for a batch that isn't done. We use an atomic counter to start
1290 : * our search at a different batch in every participant when there are
1291 : * more batches than participants.
1292 : */
1293 1258 : batchno = start_batchno =
1294 1258 : pg_atomic_fetch_add_u32(&hashtable->parallel_state->distributor, 1) %
1295 1258 : hashtable->nbatch;
1296 : do
1297 : {
1298 : uint32 hashvalue;
1299 : MinimalTuple tuple;
1300 : TupleTableSlot *slot;
1301 :
1302 3074 : if (!hashtable->batches[batchno].done)
1303 : {
1304 : SharedTuplestoreAccessor *inner_tuples;
1305 1766 : Barrier *batch_barrier =
1306 1766 : &hashtable->batches[batchno].shared->batch_barrier;
1307 :
1308 1766 : switch (BarrierAttach(batch_barrier))
1309 : {
1310 582 : case PHJ_BATCH_ELECT:
1311 :
1312 : /* One backend allocates the hash table. */
1313 582 : if (BarrierArriveAndWait(batch_barrier,
1314 : WAIT_EVENT_HASH_BATCH_ELECT))
1315 582 : ExecParallelHashTableAlloc(hashtable, batchno);
1316 : /* Fall through. */
1317 :
1318 : case PHJ_BATCH_ALLOCATE:
1319 : /* Wait for allocation to complete. */
1320 582 : BarrierArriveAndWait(batch_barrier,
1321 : WAIT_EVENT_HASH_BATCH_ALLOCATE);
1322 : /* Fall through. */
1323 :
1324 600 : case PHJ_BATCH_LOAD:
1325 : /* Start (or join in) loading tuples. */
1326 600 : ExecParallelHashTableSetCurrentBatch(hashtable, batchno);
1327 600 : inner_tuples = hashtable->batches[batchno].inner_tuples;
1328 600 : sts_begin_parallel_scan(inner_tuples);
1329 1083528 : while ((tuple = sts_parallel_scan_next(inner_tuples,
1330 : &hashvalue)))
1331 : {
1332 1082928 : ExecForceStoreMinimalTuple(tuple,
1333 : hjstate->hj_HashTupleSlot,
1334 : false);
1335 1082928 : slot = hjstate->hj_HashTupleSlot;
1336 1082928 : ExecParallelHashTableInsertCurrentBatch(hashtable, slot,
1337 : hashvalue);
1338 : }
1339 600 : sts_end_parallel_scan(inner_tuples);
1340 600 : BarrierArriveAndWait(batch_barrier,
1341 : WAIT_EVENT_HASH_BATCH_LOAD);
1342 : /* Fall through. */
1343 :
1344 866 : case PHJ_BATCH_PROBE:
1345 :
1346 : /*
1347 : * This batch is ready to probe. Return control to
1348 : * caller. We stay attached to batch_barrier so that the
1349 : * hash table stays alive until everyone's finished
1350 : * probing it, but no participant is allowed to wait at
1351 : * this barrier again (or else a deadlock could occur).
1352 : * All attached participants must eventually detach from
1353 : * the barrier and one worker must advance the phase so
1354 : * that the final phase is reached.
1355 : */
1356 866 : ExecParallelHashTableSetCurrentBatch(hashtable, batchno);
1357 866 : sts_begin_parallel_scan(hashtable->batches[batchno].outer_tuples);
1358 :
1359 866 : return true;
1360 2 : case PHJ_BATCH_SCAN:
1361 :
1362 : /*
1363 : * In principle, we could help scan for unmatched tuples,
1364 : * since that phase is already underway (the thing we
1365 : * can't do under current deadlock-avoidance rules is wait
1366 : * for others to arrive at PHJ_BATCH_SCAN, because
1367 : * PHJ_BATCH_PROBE emits tuples, but in this case we just
1368 : * got here without waiting). That is not yet done. For
1369 : * now, we just detach and go around again. We have to
1370 : * use ExecHashTableDetachBatch() because there's a small
1371 : * chance we'll be the last to detach, and then we're
1372 : * responsible for freeing memory.
1373 : */
1374 2 : ExecParallelHashTableSetCurrentBatch(hashtable, batchno);
1375 2 : hashtable->batches[batchno].done = true;
1376 2 : ExecHashTableDetachBatch(hashtable);
1377 2 : break;
1378 :
1379 898 : case PHJ_BATCH_FREE:
1380 :
1381 : /*
1382 : * Already done. Detach and go around again (if any
1383 : * remain).
1384 : */
1385 898 : BarrierDetach(batch_barrier);
1386 898 : hashtable->batches[batchno].done = true;
1387 898 : hashtable->curbatch = -1;
1388 898 : break;
1389 :
1390 0 : default:
1391 0 : elog(ERROR, "unexpected batch phase %d",
1392 : BarrierPhase(batch_barrier));
1393 : }
1394 : }
1395 2208 : batchno = (batchno + 1) % hashtable->nbatch;
1396 2208 : } while (batchno != start_batchno);
1397 :
1398 392 : return false;
1399 : }
1400 :
1401 : /*
1402 : * ExecHashJoinSaveTuple
1403 : * save a tuple to a batch file.
1404 : *
1405 : * The data recorded in the file for each tuple is its hash value,
1406 : * then the tuple in MinimalTuple format.
1407 : *
1408 : * fileptr points to a batch file in one of the hashtable arrays.
1409 : *
1410 : * The batch files (and their buffers) are allocated in the spill context
1411 : * created for the hashtable.
1412 : */
1413 : void
1414 4014274 : ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
1415 : BufFile **fileptr, HashJoinTable hashtable)
1416 : {
1417 4014274 : BufFile *file = *fileptr;
1418 :
1419 : /*
1420 : * The batch file is lazily created. If this is the first tuple written to
1421 : * this batch, the batch file is created and its buffer is allocated in
1422 : * the spillCxt context, NOT in the batchCxt.
1423 : *
1424 : * During the build phase, buffered files are created for inner batches.
1425 : * Each batch's buffered file is closed (and its buffer freed) after the
1426 : * batch is loaded into memory during the outer side scan. Therefore, it
1427 : * is necessary to allocate the batch file buffer in a memory context
1428 : * which outlives the batch itself.
1429 : *
1430 : * Also, we use spillCxt instead of hashCxt for a better accounting of the
1431 : * spilling memory consumption.
1432 : */
1433 4014274 : if (file == NULL)
1434 : {
1435 2480 : MemoryContext oldctx = MemoryContextSwitchTo(hashtable->spillCxt);
1436 :
1437 2480 : file = BufFileCreateTemp(false);
1438 2480 : *fileptr = file;
1439 :
1440 2480 : MemoryContextSwitchTo(oldctx);
1441 : }
1442 :
1443 4014274 : BufFileWrite(file, &hashvalue, sizeof(uint32));
1444 4014274 : BufFileWrite(file, tuple, tuple->t_len);
1445 4014274 : }
1446 :
1447 : /*
1448 : * ExecHashJoinGetSavedTuple
1449 : * read the next tuple from a batch file. Return NULL if no more.
1450 : *
1451 : * On success, *hashvalue is set to the tuple's hash value, and the tuple
1452 : * itself is stored in the given slot.
1453 : */
1454 : static TupleTableSlot *
1455 4016754 : ExecHashJoinGetSavedTuple(HashJoinState *hjstate,
1456 : BufFile *file,
1457 : uint32 *hashvalue,
1458 : TupleTableSlot *tupleSlot)
1459 : {
1460 : uint32 header[2];
1461 : size_t nread;
1462 : MinimalTuple tuple;
1463 :
1464 : /*
1465 : * We check for interrupts here because this is typically taken as an
1466 : * alternative code path to an ExecProcNode() call, which would include
1467 : * such a check.
1468 : */
1469 4016754 : CHECK_FOR_INTERRUPTS();
1470 :
1471 : /*
1472 : * Since both the hash value and the MinimalTuple length word are uint32,
1473 : * we can read them both in one BufFileRead() call without any type
1474 : * cheating.
1475 : */
1476 4016754 : nread = BufFileReadMaybeEOF(file, header, sizeof(header), true);
1477 4016754 : if (nread == 0) /* end of file */
1478 : {
1479 2480 : ExecClearTuple(tupleSlot);
1480 2480 : return NULL;
1481 : }
1482 4014274 : *hashvalue = header[0];
1483 4014274 : tuple = (MinimalTuple) palloc(header[1]);
1484 4014274 : tuple->t_len = header[1];
1485 4014274 : BufFileReadExact(file,
1486 : (char *) tuple + sizeof(uint32),
1487 4014274 : header[1] - sizeof(uint32));
1488 4014274 : ExecForceStoreMinimalTuple(tuple, tupleSlot, true);
1489 4014274 : return tupleSlot;
1490 : }
1491 :
1492 :
1493 : void
1494 2344 : ExecReScanHashJoin(HashJoinState *node)
1495 : {
1496 2344 : PlanState *outerPlan = outerPlanState(node);
1497 2344 : PlanState *innerPlan = innerPlanState(node);
1498 :
1499 : /*
1500 : * In a multi-batch join, we currently have to do rescans the hard way,
1501 : * primarily because batch temp files may have already been released. But
1502 : * if it's a single-batch join, and there is no parameter change for the
1503 : * inner subnode, then we can just re-use the existing hash table without
1504 : * rebuilding it.
1505 : */
1506 2344 : if (node->hj_HashTable != NULL)
1507 : {
1508 1894 : if (node->hj_HashTable->nbatch == 1 &&
1509 1894 : innerPlan->chgParam == NULL)
1510 : {
1511 : /*
1512 : * Okay to reuse the hash table; needn't rescan inner, either.
1513 : *
1514 : * However, if it's a right/right-anti/full join, we'd better
1515 : * reset the inner-tuple match flags contained in the table.
1516 : */
1517 804 : if (HJ_FILL_INNER(node))
1518 16 : ExecHashTableResetMatchFlags(node->hj_HashTable);
1519 :
1520 : /*
1521 : * Also, we need to reset our state about the emptiness of the
1522 : * outer relation, so that the new scan of the outer will update
1523 : * it correctly if it turns out to be empty this time. (There's no
1524 : * harm in clearing it now because ExecHashJoin won't need the
1525 : * info. In the other cases, where the hash table doesn't exist
1526 : * or we are destroying it, we leave this state alone because
1527 : * ExecHashJoin will need it the first time through.)
1528 : */
1529 804 : node->hj_OuterNotEmpty = false;
1530 :
1531 : /* ExecHashJoin can skip the BUILD_HASHTABLE step */
1532 804 : node->hj_JoinState = HJ_NEED_NEW_OUTER;
1533 : }
1534 : else
1535 : {
1536 : /* must destroy and rebuild hash table */
1537 1090 : HashState *hashNode = castNode(HashState, innerPlan);
1538 :
1539 : Assert(hashNode->hashtable == node->hj_HashTable);
1540 : /* accumulate stats from old hash table, if wanted */
1541 : /* (this should match ExecShutdownHash) */
1542 1090 : if (hashNode->ps.instrument && !hashNode->hinstrument)
1543 0 : hashNode->hinstrument = (HashInstrumentation *)
1544 0 : palloc0(sizeof(HashInstrumentation));
1545 1090 : if (hashNode->hinstrument)
1546 0 : ExecHashAccumInstrumentation(hashNode->hinstrument,
1547 : hashNode->hashtable);
1548 : /* for safety, be sure to clear child plan node's pointer too */
1549 1090 : hashNode->hashtable = NULL;
1550 :
1551 1090 : ExecHashTableDestroy(node->hj_HashTable);
1552 1090 : node->hj_HashTable = NULL;
1553 1090 : node->hj_JoinState = HJ_BUILD_HASHTABLE;
1554 :
1555 : /*
1556 : * if chgParam of subnode is not null then plan will be re-scanned
1557 : * by first ExecProcNode.
1558 : */
1559 1090 : if (innerPlan->chgParam == NULL)
1560 0 : ExecReScan(innerPlan);
1561 : }
1562 : }
1563 :
1564 : /* Always reset intra-tuple state */
1565 2344 : node->hj_CurHashValue = 0;
1566 2344 : node->hj_CurBucketNo = 0;
1567 2344 : node->hj_CurSkewBucketNo = INVALID_SKEW_BUCKET_NO;
1568 2344 : node->hj_CurTuple = NULL;
1569 :
1570 2344 : node->hj_MatchedOuter = false;
1571 2344 : node->hj_FirstOuterTupleSlot = NULL;
1572 :
1573 : /*
1574 : * if chgParam of subnode is not null then plan will be re-scanned by
1575 : * first ExecProcNode.
1576 : */
1577 2344 : if (outerPlan->chgParam == NULL)
1578 1596 : ExecReScan(outerPlan);
1579 2344 : }
1580 :
1581 : void
1582 27536 : ExecShutdownHashJoin(HashJoinState *node)
1583 : {
1584 27536 : if (node->hj_HashTable)
1585 : {
1586 : /*
1587 : * Detach from shared state before DSM memory goes away. This makes
1588 : * sure that we don't have any pointers into DSM memory by the time
1589 : * ExecEndHashJoin runs.
1590 : */
1591 19732 : ExecHashTableDetachBatch(node->hj_HashTable);
1592 19732 : ExecHashTableDetach(node->hj_HashTable);
1593 : }
1594 27536 : }
1595 :
1596 : static void
1597 146 : ExecParallelHashJoinPartitionOuter(HashJoinState *hjstate)
1598 : {
1599 146 : PlanState *outerState = outerPlanState(hjstate);
1600 146 : ExprContext *econtext = hjstate->js.ps.ps_ExprContext;
1601 146 : HashJoinTable hashtable = hjstate->hj_HashTable;
1602 : TupleTableSlot *slot;
1603 : uint32 hashvalue;
1604 : int i;
1605 :
1606 : Assert(hjstate->hj_FirstOuterTupleSlot == NULL);
1607 :
1608 : /* Execute outer plan, writing all tuples to shared tuplestores. */
1609 : for (;;)
1610 1200024 : {
1611 : bool isnull;
1612 :
1613 1200170 : slot = ExecProcNode(outerState);
1614 1200170 : if (TupIsNull(slot))
1615 : break;
1616 1200024 : econtext->ecxt_outertuple = slot;
1617 :
1618 1200024 : ResetExprContext(econtext);
1619 :
1620 1200024 : hashvalue = DatumGetUInt32(ExecEvalExprSwitchContext(hjstate->hj_OuterHash,
1621 : econtext,
1622 : &isnull));
1623 :
1624 1200024 : if (!isnull)
1625 : {
1626 : int batchno;
1627 : int bucketno;
1628 : bool shouldFree;
1629 1200024 : MinimalTuple mintup = ExecFetchSlotMinimalTuple(slot, &shouldFree);
1630 :
1631 1200024 : ExecHashGetBucketAndBatch(hashtable, hashvalue, &bucketno,
1632 : &batchno);
1633 1200024 : sts_puttuple(hashtable->batches[batchno].outer_tuples,
1634 : &hashvalue, mintup);
1635 :
1636 1200024 : if (shouldFree)
1637 1200024 : heap_free_minimal_tuple(mintup);
1638 : }
1639 1200024 : CHECK_FOR_INTERRUPTS();
1640 : }
1641 :
1642 : /* Make sure all outer partitions are readable by any backend. */
1643 1242 : for (i = 0; i < hashtable->nbatch; ++i)
1644 1096 : sts_end_write(hashtable->batches[i].outer_tuples);
1645 146 : }
1646 :
1647 : void
1648 120 : ExecHashJoinEstimate(HashJoinState *state, ParallelContext *pcxt)
1649 : {
1650 120 : shm_toc_estimate_chunk(&pcxt->estimator, sizeof(ParallelHashJoinState));
1651 120 : shm_toc_estimate_keys(&pcxt->estimator, 1);
1652 120 : }
1653 :
1654 : void
1655 120 : ExecHashJoinInitializeDSM(HashJoinState *state, ParallelContext *pcxt)
1656 : {
1657 120 : int plan_node_id = state->js.ps.plan->plan_node_id;
1658 : HashState *hashNode;
1659 : ParallelHashJoinState *pstate;
1660 :
1661 : /*
1662 : * Disable shared hash table mode if we failed to create a real DSM
1663 : * segment, because that means that we don't have a DSA area to work with.
1664 : */
1665 120 : if (pcxt->seg == NULL)
1666 0 : return;
1667 :
1668 120 : ExecSetExecProcNode(&state->js.ps, ExecParallelHashJoin);
1669 :
1670 : /*
1671 : * Set up the state needed to coordinate access to the shared hash
1672 : * table(s), using the plan node ID as the toc key.
1673 : */
1674 120 : pstate = shm_toc_allocate(pcxt->toc, sizeof(ParallelHashJoinState));
1675 120 : shm_toc_insert(pcxt->toc, plan_node_id, pstate);
1676 :
1677 : /*
1678 : * Set up the shared hash join state with no batches initially.
1679 : * ExecHashTableCreate() will prepare at least one later and set nbatch
1680 : * and space_allowed.
1681 : */
1682 120 : pstate->nbatch = 0;
1683 120 : pstate->space_allowed = 0;
1684 120 : pstate->batches = InvalidDsaPointer;
1685 120 : pstate->old_batches = InvalidDsaPointer;
1686 120 : pstate->nbuckets = 0;
1687 120 : pstate->growth = PHJ_GROWTH_OK;
1688 120 : pstate->chunk_work_queue = InvalidDsaPointer;
1689 120 : pg_atomic_init_u32(&pstate->distributor, 0);
1690 120 : pstate->nparticipants = pcxt->nworkers + 1;
1691 120 : pstate->total_tuples = 0;
1692 120 : LWLockInitialize(&pstate->lock,
1693 : LWTRANCHE_PARALLEL_HASH_JOIN);
1694 120 : BarrierInit(&pstate->build_barrier, 0);
1695 120 : BarrierInit(&pstate->grow_batches_barrier, 0);
1696 120 : BarrierInit(&pstate->grow_buckets_barrier, 0);
1697 :
1698 : /* Set up the space we'll use for shared temporary files. */
1699 120 : SharedFileSetInit(&pstate->fileset, pcxt->seg);
1700 :
1701 : /* Initialize the shared state in the hash node. */
1702 120 : hashNode = (HashState *) innerPlanState(state);
1703 120 : hashNode->parallel_state = pstate;
1704 : }
1705 :
1706 : /* ----------------------------------------------------------------
1707 : * ExecHashJoinReInitializeDSM
1708 : *
1709 : * Reset shared state before beginning a fresh scan.
1710 : * ----------------------------------------------------------------
1711 : */
1712 : void
1713 48 : ExecHashJoinReInitializeDSM(HashJoinState *state, ParallelContext *pcxt)
1714 : {
1715 48 : int plan_node_id = state->js.ps.plan->plan_node_id;
1716 : ParallelHashJoinState *pstate;
1717 :
1718 : /* Nothing to do if we failed to create a DSM segment. */
1719 48 : if (pcxt->seg == NULL)
1720 0 : return;
1721 :
1722 48 : pstate = shm_toc_lookup(pcxt->toc, plan_node_id, false);
1723 :
1724 : /*
1725 : * It would be possible to reuse the shared hash table in single-batch
1726 : * cases by resetting and then fast-forwarding build_barrier to
1727 : * PHJ_BUILD_FREE and batch 0's batch_barrier to PHJ_BATCH_PROBE, but
1728 : * currently shared hash tables are already freed by now (by the last
1729 : * participant to detach from the batch). We could consider keeping it
1730 : * around for single-batch joins. We'd also need to adjust
1731 : * finalize_plan() so that it doesn't record a dummy dependency for
1732 : * Parallel Hash nodes, preventing the rescan optimization. For now we
1733 : * don't try.
1734 : */
1735 :
1736 : /* Detach, freeing any remaining shared memory. */
1737 48 : if (state->hj_HashTable != NULL)
1738 : {
1739 0 : ExecHashTableDetachBatch(state->hj_HashTable);
1740 0 : ExecHashTableDetach(state->hj_HashTable);
1741 : }
1742 :
1743 : /* Clear any shared batch files. */
1744 48 : SharedFileSetDeleteAll(&pstate->fileset);
1745 :
1746 : /* Reset build_barrier to PHJ_BUILD_ELECT so we can go around again. */
1747 48 : BarrierInit(&pstate->build_barrier, 0);
1748 : }
1749 :
1750 : void
1751 308 : ExecHashJoinInitializeWorker(HashJoinState *state,
1752 : ParallelWorkerContext *pwcxt)
1753 : {
1754 : HashState *hashNode;
1755 308 : int plan_node_id = state->js.ps.plan->plan_node_id;
1756 : ParallelHashJoinState *pstate =
1757 308 : shm_toc_lookup(pwcxt->toc, plan_node_id, false);
1758 :
1759 : /* Attach to the space for shared temporary files. */
1760 308 : SharedFileSetAttach(&pstate->fileset, pwcxt->seg);
1761 :
1762 : /* Attach to the shared state in the hash node. */
1763 308 : hashNode = (HashState *) innerPlanState(state);
1764 308 : hashNode->parallel_state = pstate;
1765 :
1766 308 : ExecSetExecProcNode(&state->js.ps, ExecParallelHashJoin);
1767 308 : }
|
