granicus.if.org Git - postgresql/blob - src/backend/executor/nodeHashjoin.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * nodeHashjoin.c
   4  *        Routines to handle hash join nodes
   5  *
   6  * Portions Copyright (c) 1996-2018, 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  * PARALLELISM
  14  *
  15  * Hash joins can participate in parallel query execution in several ways.  A
  16  * parallel-oblivious hash join is one where the node is unaware that it is
  17  * part of a parallel plan.  In this case, a copy of the inner plan is used to
  18  * build a copy of the hash table in every backend, and the outer plan could
  19  * either be built from a partial or complete path, so that the results of the
  20  * hash join are correspondingly either partial or complete.  A parallel-aware
  21  * hash join is one that behaves differently, coordinating work between
  22  * backends, and appears as Parallel Hash Join in EXPLAIN output.  A Parallel
  23  * Hash Join always appears with a Parallel Hash node.
  24  *
  25  * Parallel-aware hash joins use the same per-backend state machine to track
  26  * progress through the hash join algorithm as parallel-oblivious hash joins.
  27  * In a parallel-aware hash join, there is also a shared state machine that
  28  * co-operating backends use to synchronize their local state machines and
  29  * program counters.  The shared state machine is managed with a Barrier IPC
  30  * primitive.  When all attached participants arrive at a barrier, the phase
  31  * advances and all waiting participants are released.
  32  *
  33  * When a participant begins working on a parallel hash join, it must first
  34  * figure out how much progress has already been made, because participants
  35  * don't wait for each other to begin.  For this reason there are switch
  36  * statements at key points in the code where we have to synchronize our local
  37  * state machine with the phase, and then jump to the correct part of the
  38  * algorithm so that we can get started.
  39  *
  40  * One barrier called build_barrier is used to coordinate the hashing phases.
  41  * The phase is represented by an integer which begins at zero and increments
  42  * one by one, but in the code it is referred to by symbolic names as follows:
  43  *
  44  *   PHJ_BUILD_ELECTING              -- initial state
  45  *   PHJ_BUILD_ALLOCATING            -- one sets up the batches and table 0
  46  *   PHJ_BUILD_HASHING_INNER         -- all hash the inner rel
  47  *   PHJ_BUILD_HASHING_OUTER         -- (multi-batch only) all hash the outer
  48  *   PHJ_BUILD_DONE                  -- building done, probing can begin
  49  *
  50  * While in the phase PHJ_BUILD_HASHING_INNER a separate pair of barriers may
  51  * be used repeatedly as required to coordinate expansions in the number of
  52  * batches or buckets.  Their phases are as follows:
  53  *
  54  *   PHJ_GROW_BATCHES_ELECTING       -- initial state
  55  *   PHJ_GROW_BATCHES_ALLOCATING     -- one allocates new batches
  56  *   PHJ_GROW_BATCHES_REPARTITIONING -- all repartition
  57  *   PHJ_GROW_BATCHES_FINISHING      -- one cleans up, detects skew
  58  *
  59  *   PHJ_GROW_BUCKETS_ELECTING       -- initial state
  60  *   PHJ_GROW_BUCKETS_ALLOCATING     -- one allocates new buckets
  61  *   PHJ_GROW_BUCKETS_REINSERTING    -- all insert tuples
  62  *
  63  * If the planner got the number of batches and buckets right, those won't be
  64  * necessary, but on the other hand we might finish up needing to expand the
  65  * buckets or batches multiple times while hashing the inner relation to stay
  66  * within our memory budget and load factor target.  For that reason it's a
  67  * separate pair of barriers using circular phases.
  68  *
  69  * The PHJ_BUILD_HASHING_OUTER phase is required only for multi-batch joins,
  70  * because we need to divide the outer relation into batches up front in order
  71  * to be able to process batches entirely independently.  In contrast, the
  72  * parallel-oblivious algorithm simply throws tuples 'forward' to 'later'
  73  * batches whenever it encounters them while scanning and probing, which it
  74  * can do because it processes batches in serial order.
  75  *
  76  * Once PHJ_BUILD_DONE is reached, backends then split up and process
  77  * different batches, or gang up and work together on probing batches if there
  78  * aren't enough to go around.  For each batch there is a separate barrier
  79  * with the following phases:
  80  *
  81  *  PHJ_BATCH_ELECTING       -- initial state
  82  *  PHJ_BATCH_ALLOCATING     -- one allocates buckets
  83  *  PHJ_BATCH_LOADING        -- all load the hash table from disk
  84  *  PHJ_BATCH_PROBING        -- all probe
  85  *  PHJ_BATCH_DONE           -- end
  86  *
  87  * Batch 0 is a special case, because it starts out in phase
  88  * PHJ_BATCH_PROBING; populating batch 0's hash table is done during
  89  * PHJ_BUILD_HASHING_INNER so we can skip loading.
  90  *
  91  * Initially we try to plan for a single-batch hash join using the combined
  92  * work_mem of all participants to create a large shared hash table.  If that
  93  * turns out either at planning or execution time to be impossible then we
  94  * fall back to regular work_mem sized hash tables.
  95  *
  96  * To avoid deadlocks, we never wait for any barrier unless it is known that
  97  * all other backends attached to it are actively executing the node or have
  98  * already arrived.  Practically, that means that we never return a tuple
  99  * while attached to a barrier, unless the barrier has reached its final
 100  * state.  In the slightly special case of the per-batch barrier, we return
 101  * tuples while in PHJ_BATCH_PROBING phase, but that's OK because we use
 102  * BarrierArriveAndDetach() to advance it to PHJ_BATCH_DONE without waiting.
 103  *
 104  *-------------------------------------------------------------------------
 105  */
 106
 107 #include "postgres.h"
 108
 109 #include "access/htup_details.h"
 110 #include "access/parallel.h"
 111 #include "executor/executor.h"
 112 #include "executor/hashjoin.h"
 113 #include "executor/nodeHash.h"
 114 #include "executor/nodeHashjoin.h"
 115 #include "miscadmin.h"
 116 #include "pgstat.h"
 117 #include "utils/memutils.h"
 118 #include "utils/sharedtuplestore.h"
 119
 120
 121 /*
 122  * States of the ExecHashJoin state machine
 123  */
 124 #define HJ_BUILD_HASHTABLE              1
 125 #define HJ_NEED_NEW_OUTER               2
 126 #define HJ_SCAN_BUCKET                  3
 127 #define HJ_FILL_OUTER_TUPLE             4
 128 #define HJ_FILL_INNER_TUPLES    5
 129 #define HJ_NEED_NEW_BATCH               6
 130
 131 /* Returns true if doing null-fill on outer relation */
 132 #define HJ_FILL_OUTER(hjstate)  ((hjstate)->hj_NullInnerTupleSlot != NULL)
 133 /* Returns true if doing null-fill on inner relation */
 134 #define HJ_FILL_INNER(hjstate)  ((hjstate)->hj_NullOuterTupleSlot != NULL)
 135
 136 static TupleTableSlot *ExecHashJoinOuterGetTuple(PlanState *outerNode,
 137                                                   HashJoinState *hjstate,
 138                                                   uint32 *hashvalue);
 139 static TupleTableSlot *ExecParallelHashJoinOuterGetTuple(PlanState *outerNode,
 140                                                                   HashJoinState *hjstate,
 141                                                                   uint32 *hashvalue);
 142 static TupleTableSlot *ExecHashJoinGetSavedTuple(HashJoinState *hjstate,
 143                                                   BufFile *file,
 144                                                   uint32 *hashvalue,
 145                                                   TupleTableSlot *tupleSlot);
 146 static bool ExecHashJoinNewBatch(HashJoinState *hjstate);
 147 static bool ExecParallelHashJoinNewBatch(HashJoinState *hjstate);
 148 static void ExecParallelHashJoinPartitionOuter(HashJoinState *node);
 149
 150
 151 /* ----------------------------------------------------------------
 152  *              ExecHashJoinImpl
 153  *
 154  *              This function implements the Hybrid Hashjoin algorithm.  It is marked
 155  *              with an always-inline attribute so that ExecHashJoin() and
 156  *              ExecParallelHashJoin() can inline it.  Compilers that respect the
 157  *              attribute should create versions specialized for parallel == true and
 158  *              parallel == false with unnecessary branches removed.
 159  *
 160  *              Note: the relation we build hash table on is the "inner"
 161  *                        the other one is "outer".
 162  * ----------------------------------------------------------------
 163  */
 164 static pg_attribute_always_inline TupleTableSlot *
 165 ExecHashJoinImpl(PlanState *pstate, bool parallel)
 166 {
 167         HashJoinState *node = castNode(HashJoinState, pstate);
 168         PlanState  *outerNode;
 169         HashState  *hashNode;
 170         ExprState  *joinqual;
 171         ExprState  *otherqual;
 172         ExprContext *econtext;
 173         HashJoinTable hashtable;
 174         TupleTableSlot *outerTupleSlot;
 175         uint32          hashvalue;
 176         int                     batchno;
 177         ParallelHashJoinState *parallel_state;
 178
 179         /*
 180          * get information from HashJoin node
 181          */
 182         joinqual = node->js.joinqual;
 183         otherqual = node->js.ps.qual;
 184         hashNode = (HashState *) innerPlanState(node);
 185         outerNode = outerPlanState(node);
 186         hashtable = node->hj_HashTable;
 187         econtext = node->js.ps.ps_ExprContext;
 188         parallel_state = hashNode->parallel_state;
 189
 190         /*
 191          * Reset per-tuple memory context to free any expression evaluation
 192          * storage allocated in the previous tuple cycle.
 193          */
 194         ResetExprContext(econtext);
 195
 196         /*
 197          * run the hash join state machine
 198          */
 199         for (;;)
 200         {
 201                 /*
 202                  * It's possible to iterate this loop many times before returning a
 203                  * tuple, in some pathological cases such as needing to move much of
 204                  * the current batch to a later batch.  So let's check for interrupts
 205                  * each time through.
 206                  */
 207                 CHECK_FOR_INTERRUPTS();
 208
 209                 switch (node->hj_JoinState)
 210                 {
 211                         case HJ_BUILD_HASHTABLE:
 212
 213                                 /*
 214                                  * First time through: build hash table for inner relation.
 215                                  */
 216                                 Assert(hashtable == NULL);
 217
 218                                 /*
 219                                  * If the outer relation is completely empty, and it's not
 220                                  * right/full join, we can quit without building the hash
 221                                  * table.  However, for an inner join it is only a win to
 222                                  * check this when the outer relation's startup cost is less
 223                                  * than the projected cost of building the hash table.
 224                                  * Otherwise it's best to build the hash table first and see
 225                                  * if the inner relation is empty.  (When it's a left join, we
 226                                  * should always make this check, since we aren't going to be
 227                                  * able to skip the join on the strength of an empty inner
 228                                  * relation anyway.)
 229                                  *
 230                                  * If we are rescanning the join, we make use of information
 231                                  * gained on the previous scan: don't bother to try the
 232                                  * prefetch if the previous scan found the outer relation
 233                                  * nonempty. This is not 100% reliable since with new
 234                                  * parameters the outer relation might yield different
 235                                  * results, but it's a good heuristic.
 236                                  *
 237                                  * The only way to make the check is to try to fetch a tuple
 238                                  * from the outer plan node.  If we succeed, we have to stash
 239                                  * it away for later consumption by ExecHashJoinOuterGetTuple.
 240                                  */
 241                                 if (HJ_FILL_INNER(node))
 242                                 {
 243                                         /* no chance to not build the hash table */
 244                                         node->hj_FirstOuterTupleSlot = NULL;
 245                                 }
 246                                 else if (parallel)
 247                                 {
 248                                         /*
 249                                          * The empty-outer optimization is not implemented for
 250                                          * shared hash tables, because no one participant can
 251                                          * determine that there are no outer tuples, and it's not
 252                                          * yet clear that it's worth the synchronization overhead
 253                                          * of reaching consensus to figure that out.  So we have
 254                                          * to build the hash table.
 255                                          */
 256                                         node->hj_FirstOuterTupleSlot = NULL;
 257                                 }
 258                                 else if (HJ_FILL_OUTER(node) ||
 259                                                  (outerNode->plan->startup_cost < hashNode->ps.plan->total_cost &&
 260                                                   !node->hj_OuterNotEmpty))
 261                                 {
 262                                         node->hj_FirstOuterTupleSlot = ExecProcNode(outerNode);
 263                                         if (TupIsNull(node->hj_FirstOuterTupleSlot))
 264                                         {
 265                                                 node->hj_OuterNotEmpty = false;
 266                                                 return NULL;
 267                                         }
 268                                         else
 269                                                 node->hj_OuterNotEmpty = true;
 270                                 }
 271                                 else
 272                                         node->hj_FirstOuterTupleSlot = NULL;
 273
 274                                 /*
 275                                  * Create the hash table.  If using Parallel Hash, then
 276                                  * whoever gets here first will create the hash table and any
 277                                  * later arrivals will merely attach to it.
 278                                  */
 279                                 hashtable = ExecHashTableCreate(hashNode,
 280                                                                                                 node->hj_HashOperators,
 281                                                                                                 HJ_FILL_INNER(node));
 282                                 node->hj_HashTable = hashtable;
 283
 284                                 /*
 285                                  * Execute the Hash node, to build the hash table.  If using
 286                                  * Parallel Hash, then we'll try to help hashing unless we
 287                                  * arrived too late.
 288                                  */
 289                                 hashNode->hashtable = hashtable;
 290                                 (void) MultiExecProcNode((PlanState *) hashNode);
 291
 292                                 /*
 293                                  * If the inner relation is completely empty, and we're not
 294                                  * doing a left outer join, we can quit without scanning the
 295                                  * outer relation.
 296                                  */
 297                                 if (hashtable->totalTuples == 0 && !HJ_FILL_OUTER(node))
 298                                         return NULL;
 299
 300                                 /*
 301                                  * need to remember whether nbatch has increased since we
 302                                  * began scanning the outer relation
 303                                  */
 304                                 hashtable->nbatch_outstart = hashtable->nbatch;
 305
 306                                 /*
 307                                  * Reset OuterNotEmpty for scan.  (It's OK if we fetched a
 308                                  * tuple above, because ExecHashJoinOuterGetTuple will
 309                                  * immediately set it again.)
 310                                  */
 311                                 node->hj_OuterNotEmpty = false;
 312
 313                                 if (parallel)
 314                                 {
 315                                         Barrier    *build_barrier;
 316
 317                                         build_barrier = &parallel_state->build_barrier;
 318                                         Assert(BarrierPhase(build_barrier) == PHJ_BUILD_HASHING_OUTER ||
 319                                                    BarrierPhase(build_barrier) == PHJ_BUILD_DONE);
 320                                         if (BarrierPhase(build_barrier) == PHJ_BUILD_HASHING_OUTER)
 321                                         {
 322                                                 /*
 323                                                  * If multi-batch, we need to hash the outer relation
 324                                                  * up front.
 325                                                  */
 326                                                 if (hashtable->nbatch > 1)
 327                                                         ExecParallelHashJoinPartitionOuter(node);
 328                                                 BarrierArriveAndWait(build_barrier,
 329                                                                                          WAIT_EVENT_HASH_BUILD_HASHING_OUTER);
 330                                         }
 331                                         Assert(BarrierPhase(build_barrier) == PHJ_BUILD_DONE);
 332
 333                                         /* Each backend should now select a batch to work on. */
 334                                         hashtable->curbatch = -1;
 335                                         node->hj_JoinState = HJ_NEED_NEW_BATCH;
 336
 337                                         continue;
 338                                 }
 339                                 else
 340                                         node->hj_JoinState = HJ_NEED_NEW_OUTER;
 341
 342                                 /* FALL THRU */
 343
 344                         case HJ_NEED_NEW_OUTER:
 345
 346                                 /*
 347                                  * We don't have an outer tuple, try to get the next one
 348                                  */
 349                                 if (parallel)
 350                                         outerTupleSlot =
 351                                                 ExecParallelHashJoinOuterGetTuple(outerNode, node,
 352                                                                                                                   &hashvalue);
 353                                 else
 354                                         outerTupleSlot =
 355                                                 ExecHashJoinOuterGetTuple(outerNode, node, &hashvalue);
 356
 357                                 if (TupIsNull(outerTupleSlot))
 358                                 {
 359                                         /* end of batch, or maybe whole join */
 360                                         if (HJ_FILL_INNER(node))
 361                                         {
 362                                                 /* set up to scan for unmatched inner tuples */
 363                                                 ExecPrepHashTableForUnmatched(node);
 364                                                 node->hj_JoinState = HJ_FILL_INNER_TUPLES;
 365                                         }
 366                                         else
 367                                                 node->hj_JoinState = HJ_NEED_NEW_BATCH;
 368                                         continue;
 369                                 }
 370
 371                                 econtext->ecxt_outertuple = outerTupleSlot;
 372                                 node->hj_MatchedOuter = false;
 373
 374                                 /*
 375                                  * Find the corresponding bucket for this tuple in the main
 376                                  * hash table or skew hash table.
 377                                  */
 378                                 node->hj_CurHashValue = hashvalue;
 379                                 ExecHashGetBucketAndBatch(hashtable, hashvalue,
 380                                                                                   &node->hj_CurBucketNo, &batchno);
 381                                 node->hj_CurSkewBucketNo = ExecHashGetSkewBucket(hashtable,
 382                                                                                                                                  hashvalue);
 383                                 node->hj_CurTuple = NULL;
 384
 385                                 /*
 386                                  * The tuple might not belong to the current batch (where
 387                                  * "current batch" includes the skew buckets if any).
 388                                  */
 389                                 if (batchno != hashtable->curbatch &&
 390                                         node->hj_CurSkewBucketNo == INVALID_SKEW_BUCKET_NO)
 391                                 {
 392                                         /*
 393                                          * Need to postpone this outer tuple to a later batch.
 394                                          * Save it in the corresponding outer-batch file.
 395                                          */
 396                                         Assert(parallel_state == NULL);
 397                                         Assert(batchno > hashtable->curbatch);
 398                                         ExecHashJoinSaveTuple(ExecFetchSlotMinimalTuple(outerTupleSlot),
 399                                                                                   hashvalue,
 400                                                                                   &hashtable->outerBatchFile[batchno]);
 401
 402                                         /* Loop around, staying in HJ_NEED_NEW_OUTER state */
 403                                         continue;
 404                                 }
 405
 406                                 /* OK, let's scan the bucket for matches */
 407                                 node->hj_JoinState = HJ_SCAN_BUCKET;
 408
 409                                 /* FALL THRU */
 410
 411                         case HJ_SCAN_BUCKET:
 412
 413                                 /*
 414                                  * Scan the selected hash bucket for matches to current outer
 415                                  */
 416                                 if (parallel)
 417                                 {
 418                                         if (!ExecParallelScanHashBucket(node, econtext))
 419                                         {
 420                                                 /* out of matches; check for possible outer-join fill */
 421                                                 node->hj_JoinState = HJ_FILL_OUTER_TUPLE;
 422                                                 continue;
 423                                         }
 424                                 }
 425                                 else
 426                                 {
 427                                         if (!ExecScanHashBucket(node, econtext))
 428                                         {
 429                                                 /* out of matches; check for possible outer-join fill */
 430                                                 node->hj_JoinState = HJ_FILL_OUTER_TUPLE;
 431                                                 continue;
 432                                         }
 433                                 }
 434
 435                                 /*
 436                                  * We've got a match, but still need to test non-hashed quals.
 437                                  * ExecScanHashBucket already set up all the state needed to
 438                                  * call ExecQual.
 439                                  *
 440                                  * If we pass the qual, then save state for next call and have
 441                                  * ExecProject form the projection, store it in the tuple
 442                                  * table, and return the slot.
 443                                  *
 444                                  * Only the joinquals determine tuple match status, but all
 445                                  * quals must pass to actually return the tuple.
 446                                  */
 447                                 if (joinqual == NULL || ExecQual(joinqual, econtext))
 448                                 {
 449                                         node->hj_MatchedOuter = true;
 450                                         HeapTupleHeaderSetMatch(HJTUPLE_MINTUPLE(node->hj_CurTuple));
 451
 452                                         /* In an antijoin, we never return a matched tuple */
 453                                         if (node->js.jointype == JOIN_ANTI)
 454                                         {
 455                                                 node->hj_JoinState = HJ_NEED_NEW_OUTER;
 456                                                 continue;
 457                                         }
 458
 459                                         /*
 460                                          * If we only need to join to the first matching inner
 461                                          * tuple, then consider returning this one, but after that
 462                                          * continue with next outer tuple.
 463                                          */
 464                                         if (node->js.single_match)
 465                                                 node->hj_JoinState = HJ_NEED_NEW_OUTER;
 466
 467                                         if (otherqual == NULL || ExecQual(otherqual, econtext))
 468                                                 return ExecProject(node->js.ps.ps_ProjInfo);
 469                                         else
 470                                                 InstrCountFiltered2(node, 1);
 471                                 }
 472                                 else
 473                                         InstrCountFiltered1(node, 1);
 474                                 break;
 475
 476                         case HJ_FILL_OUTER_TUPLE:
 477
 478                                 /*
 479                                  * The current outer tuple has run out of matches, so check
 480                                  * whether to emit a dummy outer-join tuple.  Whether we emit
 481                                  * one or not, the next state is NEED_NEW_OUTER.
 482                                  */
 483                                 node->hj_JoinState = HJ_NEED_NEW_OUTER;
 484
 485                                 if (!node->hj_MatchedOuter &&
 486                                         HJ_FILL_OUTER(node))
 487                                 {
 488                                         /*
 489                                          * Generate a fake join tuple with nulls for the inner
 490                                          * tuple, and return it if it passes the non-join quals.
 491                                          */
 492                                         econtext->ecxt_innertuple = node->hj_NullInnerTupleSlot;
 493
 494                                         if (otherqual == NULL || ExecQual(otherqual, econtext))
 495                                                 return ExecProject(node->js.ps.ps_ProjInfo);
 496                                         else
 497                                                 InstrCountFiltered2(node, 1);
 498                                 }
 499                                 break;
 500
 501                         case HJ_FILL_INNER_TUPLES:
 502
 503                                 /*
 504                                  * We have finished a batch, but we are doing right/full join,
 505                                  * so any unmatched inner tuples in the hashtable have to be
 506                                  * emitted before we continue to the next batch.
 507                                  */
 508                                 if (!ExecScanHashTableForUnmatched(node, econtext))
 509                                 {
 510                                         /* no more unmatched tuples */
 511                                         node->hj_JoinState = HJ_NEED_NEW_BATCH;
 512                                         continue;
 513                                 }
 514
 515                                 /*
 516                                  * Generate a fake join tuple with nulls for the outer tuple,
 517                                  * and return it if it passes the non-join quals.
 518                                  */
 519                                 econtext->ecxt_outertuple = node->hj_NullOuterTupleSlot;
 520
 521                                 if (otherqual == NULL || ExecQual(otherqual, econtext))
 522                                         return ExecProject(node->js.ps.ps_ProjInfo);
 523                                 else
 524                                         InstrCountFiltered2(node, 1);
 525                                 break;
 526
 527                         case HJ_NEED_NEW_BATCH:
 528
 529                                 /*
 530                                  * Try to advance to next batch.  Done if there are no more.
 531                                  */
 532                                 if (parallel)
 533                                 {
 534                                         if (!ExecParallelHashJoinNewBatch(node))
 535                                                 return NULL;    /* end of parallel-aware join */
 536                                 }
 537                                 else
 538                                 {
 539                                         if (!ExecHashJoinNewBatch(node))
 540                                                 return NULL;    /* end of parallel-oblivious join */
 541                                 }
 542                                 node->hj_JoinState = HJ_NEED_NEW_OUTER;
 543                                 break;
 544
 545                         default:
 546                                 elog(ERROR, "unrecognized hashjoin state: %d",
 547                                          (int) node->hj_JoinState);
 548                 }
 549         }
 550 }
 551
 552 /* ----------------------------------------------------------------
 553  *              ExecHashJoin
 554  *
 555  *              Parallel-oblivious version.
 556  * ----------------------------------------------------------------
 557  */
 558 static TupleTableSlot *                 /* return: a tuple or NULL */
 559 ExecHashJoin(PlanState *pstate)
 560 {
 561         /*
 562          * On sufficiently smart compilers this should be inlined with the
 563          * parallel-aware branches removed.
 564          */
 565         return ExecHashJoinImpl(pstate, false);
 566 }
 567
 568 /* ----------------------------------------------------------------
 569  *              ExecParallelHashJoin
 570  *
 571  *              Parallel-aware version.
 572  * ----------------------------------------------------------------
 573  */
 574 static TupleTableSlot *                 /* return: a tuple or NULL */
 575 ExecParallelHashJoin(PlanState *pstate)
 576 {
 577         /*
 578          * On sufficiently smart compilers this should be inlined with the
 579          * parallel-oblivious branches removed.
 580          */
 581         return ExecHashJoinImpl(pstate, true);
 582 }
 583
 584 /* ----------------------------------------------------------------
 585  *              ExecInitHashJoin
 586  *
 587  *              Init routine for HashJoin node.
 588  * ----------------------------------------------------------------
 589  */
 590 HashJoinState *
 591 ExecInitHashJoin(HashJoin *node, EState *estate, int eflags)
 592 {
 593         HashJoinState *hjstate;
 594         Plan       *outerNode;
 595         Hash       *hashNode;
 596         List       *lclauses;
 597         List       *rclauses;
 598         List       *hoperators;
 599         TupleDesc       outerDesc,
 600                                 innerDesc;
 601         ListCell   *l;
 602
 603         /* check for unsupported flags */
 604         Assert(!(eflags & (EXEC_FLAG_BACKWARD | EXEC_FLAG_MARK)));
 605
 606         /*
 607          * create state structure
 608          */
 609         hjstate = makeNode(HashJoinState);
 610         hjstate->js.ps.plan = (Plan *) node;
 611         hjstate->js.ps.state = estate;
 612
 613         /*
 614          * See ExecHashJoinInitializeDSM() and ExecHashJoinInitializeWorker()
 615          * where this function may be replaced with a parallel version, if we
 616          * managed to launch a parallel query.
 617          */
 618         hjstate->js.ps.ExecProcNode = ExecHashJoin;
 619         hjstate->js.jointype = node->join.jointype;
 620
 621         /*
 622          * Miscellaneous initialization
 623          *
 624          * create expression context for node
 625          */
 626         ExecAssignExprContext(estate, &hjstate->js.ps);
 627
 628         /*
 629          * initialize child nodes
 630          *
 631          * Note: we could suppress the REWIND flag for the inner input, which
 632          * would amount to betting that the hash will be a single batch.  Not
 633          * clear if this would be a win or not.
 634          */
 635         outerNode = outerPlan(node);
 636         hashNode = (Hash *) innerPlan(node);
 637
 638         outerPlanState(hjstate) = ExecInitNode(outerNode, estate, eflags);
 639         outerDesc = ExecGetResultType(outerPlanState(hjstate));
 640         innerPlanState(hjstate) = ExecInitNode((Plan *) hashNode, estate, eflags);
 641         innerDesc = ExecGetResultType(innerPlanState(hjstate));
 642
 643         /*
 644          * Initialize result slot, type and projection.
 645          */
 646         ExecInitResultTupleSlotTL(estate, &hjstate->js.ps);
 647         ExecAssignProjectionInfo(&hjstate->js.ps, NULL);
 648
 649         /*
 650          * tuple table initialization
 651          */
 652         hjstate->hj_OuterTupleSlot = ExecInitExtraTupleSlot(estate, outerDesc);
 653
 654         /*
 655          * detect whether we need only consider the first matching inner tuple
 656          */
 657         hjstate->js.single_match = (node->join.inner_unique ||
 658                                                                 node->join.jointype == JOIN_SEMI);
 659
 660         /* set up null tuples for outer joins, if needed */
 661         switch (node->join.jointype)
 662         {
 663                 case JOIN_INNER:
 664                 case JOIN_SEMI:
 665                         break;
 666                 case JOIN_LEFT:
 667                 case JOIN_ANTI:
 668                         hjstate->hj_NullInnerTupleSlot =
 669                                 ExecInitNullTupleSlot(estate, innerDesc);
 670                         break;
 671                 case JOIN_RIGHT:
 672                         hjstate->hj_NullOuterTupleSlot =
 673                                 ExecInitNullTupleSlot(estate, outerDesc);
 674                         break;
 675                 case JOIN_FULL:
 676                         hjstate->hj_NullOuterTupleSlot =
 677                                 ExecInitNullTupleSlot(estate, outerDesc);
 678                         hjstate->hj_NullInnerTupleSlot =
 679                                 ExecInitNullTupleSlot(estate, innerDesc);
 680                         break;
 681                 default:
 682                         elog(ERROR, "unrecognized join type: %d",
 683                                  (int) node->join.jointype);
 684         }
 685
 686         /*
 687          * now for some voodoo.  our temporary tuple slot is actually the result
 688          * tuple slot of the Hash node (which is our inner plan).  we can do this
 689          * because Hash nodes don't return tuples via ExecProcNode() -- instead
 690          * the hash join node uses ExecScanHashBucket() to get at the contents of
 691          * the hash table.  -cim 6/9/91
 692          */
 693         {
 694                 HashState  *hashstate = (HashState *) innerPlanState(hjstate);
 695                 TupleTableSlot *slot = hashstate->ps.ps_ResultTupleSlot;
 696
 697                 hjstate->hj_HashTupleSlot = slot;
 698         }
 699
 700         /*
 701          * initialize child expressions
 702          */
 703         hjstate->js.ps.qual =
 704                 ExecInitQual(node->join.plan.qual, (PlanState *) hjstate);
 705         hjstate->js.joinqual =
 706                 ExecInitQual(node->join.joinqual, (PlanState *) hjstate);
 707         hjstate->hashclauses =
 708                 ExecInitQual(node->hashclauses, (PlanState *) hjstate);
 709
 710         /*
 711          * initialize hash-specific info
 712          */
 713         hjstate->hj_HashTable = NULL;
 714         hjstate->hj_FirstOuterTupleSlot = NULL;
 715
 716         hjstate->hj_CurHashValue = 0;
 717         hjstate->hj_CurBucketNo = 0;
 718         hjstate->hj_CurSkewBucketNo = INVALID_SKEW_BUCKET_NO;
 719         hjstate->hj_CurTuple = NULL;
 720
 721         /*
 722          * Deconstruct the hash clauses into outer and inner argument values, so
 723          * that we can evaluate those subexpressions separately.  Also make a list
 724          * of the hash operator OIDs, in preparation for looking up the hash
 725          * functions to use.
 726          */
 727         lclauses = NIL;
 728         rclauses = NIL;
 729         hoperators = NIL;
 730         foreach(l, node->hashclauses)
 731         {
 732                 OpExpr     *hclause = lfirst_node(OpExpr, l);
 733
 734                 lclauses = lappend(lclauses, ExecInitExpr(linitial(hclause->args),
 735                                                                                                   (PlanState *) hjstate));
 736                 rclauses = lappend(rclauses, ExecInitExpr(lsecond(hclause->args),
 737                                                                                                   (PlanState *) hjstate));
 738                 hoperators = lappend_oid(hoperators, hclause->opno);
 739         }
 740         hjstate->hj_OuterHashKeys = lclauses;
 741         hjstate->hj_InnerHashKeys = rclauses;
 742         hjstate->hj_HashOperators = hoperators;
 743         /* child Hash node needs to evaluate inner hash keys, too */
 744         ((HashState *) innerPlanState(hjstate))->hashkeys = rclauses;
 745
 746         hjstate->hj_JoinState = HJ_BUILD_HASHTABLE;
 747         hjstate->hj_MatchedOuter = false;
 748         hjstate->hj_OuterNotEmpty = false;
 749
 750         return hjstate;
 751 }
 752
 753 /* ----------------------------------------------------------------
 754  *              ExecEndHashJoin
 755  *
 756  *              clean up routine for HashJoin node
 757  * ----------------------------------------------------------------
 758  */
 759 void
 760 ExecEndHashJoin(HashJoinState *node)
 761 {
 762         /*
 763          * Free hash table
 764          */
 765         if (node->hj_HashTable)
 766         {
 767                 ExecHashTableDestroy(node->hj_HashTable);
 768                 node->hj_HashTable = NULL;
 769         }
 770
 771         /*
 772          * Free the exprcontext
 773          */
 774         ExecFreeExprContext(&node->js.ps);
 775
 776         /*
 777          * clean out the tuple table
 778          */
 779         ExecClearTuple(node->js.ps.ps_ResultTupleSlot);
 780         ExecClearTuple(node->hj_OuterTupleSlot);
 781         ExecClearTuple(node->hj_HashTupleSlot);
 782
 783         /*
 784          * clean up subtrees
 785          */
 786         ExecEndNode(outerPlanState(node));
 787         ExecEndNode(innerPlanState(node));
 788 }
 789
 790 /*
 791  * ExecHashJoinOuterGetTuple
 792  *
 793  *              get the next outer tuple for a parallel oblivious hashjoin: either by
 794  *              executing the outer plan node in the first pass, or from the temp
 795  *              files for the hashjoin batches.
 796  *
 797  * Returns a null slot if no more outer tuples (within the current batch).
 798  *
 799  * On success, the tuple's hash value is stored at *hashvalue --- this is
 800  * either originally computed, or re-read from the temp file.
 801  */
 802 static TupleTableSlot *
 803 ExecHashJoinOuterGetTuple(PlanState *outerNode,
 804                                                   HashJoinState *hjstate,
 805                                                   uint32 *hashvalue)
 806 {
 807         HashJoinTable hashtable = hjstate->hj_HashTable;
 808         int                     curbatch = hashtable->curbatch;
 809         TupleTableSlot *slot;
 810
 811         if (curbatch == 0)                      /* if it is the first pass */
 812         {
 813                 /*
 814                  * Check to see if first outer tuple was already fetched by
 815                  * ExecHashJoin() and not used yet.
 816                  */
 817                 slot = hjstate->hj_FirstOuterTupleSlot;
 818                 if (!TupIsNull(slot))
 819                         hjstate->hj_FirstOuterTupleSlot = NULL;
 820                 else
 821                         slot = ExecProcNode(outerNode);
 822
 823                 while (!TupIsNull(slot))
 824                 {
 825                         /*
 826                          * We have to compute the tuple's hash value.
 827                          */
 828                         ExprContext *econtext = hjstate->js.ps.ps_ExprContext;
 829
 830                         econtext->ecxt_outertuple = slot;
 831                         if (ExecHashGetHashValue(hashtable, econtext,
 832                                                                          hjstate->hj_OuterHashKeys,
 833                                                                          true,  /* outer tuple */
 834                                                                          HJ_FILL_OUTER(hjstate),
 835                                                                          hashvalue))
 836                         {
 837                                 /* remember outer relation is not empty for possible rescan */
 838                                 hjstate->hj_OuterNotEmpty = true;
 839
 840                                 return slot;
 841                         }
 842
 843                         /*
 844                          * That tuple couldn't match because of a NULL, so discard it and
 845                          * continue with the next one.
 846                          */
 847                         slot = ExecProcNode(outerNode);
 848                 }
 849         }
 850         else if (curbatch < hashtable->nbatch)
 851         {
 852                 BufFile    *file = hashtable->outerBatchFile[curbatch];
 853
 854                 /*
 855                  * In outer-join cases, we could get here even though the batch file
 856                  * is empty.
 857                  */
 858                 if (file == NULL)
 859                         return NULL;
 860
 861                 slot = ExecHashJoinGetSavedTuple(hjstate,
 862                                                                                  file,
 863                                                                                  hashvalue,
 864                                                                                  hjstate->hj_OuterTupleSlot);
 865                 if (!TupIsNull(slot))
 866                         return slot;
 867         }
 868
 869         /* End of this batch */
 870         return NULL;
 871 }
 872
 873 /*
 874  * ExecHashJoinOuterGetTuple variant for the parallel case.
 875  */
 876 static TupleTableSlot *
 877 ExecParallelHashJoinOuterGetTuple(PlanState *outerNode,
 878                                                                   HashJoinState *hjstate,
 879                                                                   uint32 *hashvalue)
 880 {
 881         HashJoinTable hashtable = hjstate->hj_HashTable;
 882         int                     curbatch = hashtable->curbatch;
 883         TupleTableSlot *slot;
 884
 885         /*
 886          * In the Parallel Hash case we only run the outer plan directly for
 887          * single-batch hash joins.  Otherwise we have to go to batch files, even
 888          * for batch 0.
 889          */
 890         if (curbatch == 0 && hashtable->nbatch == 1)
 891         {
 892                 slot = ExecProcNode(outerNode);
 893
 894                 while (!TupIsNull(slot))
 895                 {
 896                         ExprContext *econtext = hjstate->js.ps.ps_ExprContext;
 897
 898                         econtext->ecxt_outertuple = slot;
 899                         if (ExecHashGetHashValue(hashtable, econtext,
 900                                                                          hjstate->hj_OuterHashKeys,
 901                                                                          true,  /* outer tuple */
 902                                                                          HJ_FILL_OUTER(hjstate),
 903                                                                          hashvalue))
 904                                 return slot;
 905
 906                         /*
 907                          * That tuple couldn't match because of a NULL, so discard it and
 908                          * continue with the next one.
 909                          */
 910                         slot = ExecProcNode(outerNode);
 911                 }
 912         }
 913         else if (curbatch < hashtable->nbatch)
 914         {
 915                 MinimalTuple tuple;
 916
 917                 tuple = sts_parallel_scan_next(hashtable->batches[curbatch].outer_tuples,
 918                                                                            hashvalue);
 919                 if (tuple != NULL)
 920                 {
 921                         slot = ExecStoreMinimalTuple(tuple,
 922                                                                                  hjstate->hj_OuterTupleSlot,
 923                                                                                  false);
 924                         return slot;
 925                 }
 926                 else
 927                         ExecClearTuple(hjstate->hj_OuterTupleSlot);
 928         }
 929
 930         /* End of this batch */
 931         return NULL;
 932 }
 933
 934 /*
 935  * ExecHashJoinNewBatch
 936  *              switch to a new hashjoin batch
 937  *
 938  * Returns true if successful, false if there are no more batches.
 939  */
 940 static bool
 941 ExecHashJoinNewBatch(HashJoinState *hjstate)
 942 {
 943         HashJoinTable hashtable = hjstate->hj_HashTable;
 944         int                     nbatch;
 945         int                     curbatch;
 946         BufFile    *innerFile;
 947         TupleTableSlot *slot;
 948         uint32          hashvalue;
 949
 950         nbatch = hashtable->nbatch;
 951         curbatch = hashtable->curbatch;
 952
 953         if (curbatch > 0)
 954         {
 955                 /*
 956                  * We no longer need the previous outer batch file; close it right
 957                  * away to free disk space.
 958                  */
 959                 if (hashtable->outerBatchFile[curbatch])
 960                         BufFileClose(hashtable->outerBatchFile[curbatch]);
 961                 hashtable->outerBatchFile[curbatch] = NULL;
 962         }
 963         else                                            /* we just finished the first batch */
 964         {
 965                 /*
 966                  * Reset some of the skew optimization state variables, since we no
 967                  * longer need to consider skew tuples after the first batch. The
 968                  * memory context reset we are about to do will release the skew
 969                  * hashtable itself.
 970                  */
 971                 hashtable->skewEnabled = false;
 972                 hashtable->skewBucket = NULL;
 973                 hashtable->skewBucketNums = NULL;
 974                 hashtable->nSkewBuckets = 0;
 975                 hashtable->spaceUsedSkew = 0;
 976         }
 977
 978         /*
 979          * We can always skip over any batches that are completely empty on both
 980          * sides.  We can sometimes skip over batches that are empty on only one
 981          * side, but there are exceptions:
 982          *
 983          * 1. In a left/full outer join, we have to process outer batches even if
 984          * the inner batch is empty.  Similarly, in a right/full outer join, we
 985          * have to process inner batches even if the outer batch is empty.
 986          *
 987          * 2. If we have increased nbatch since the initial estimate, we have to
 988          * scan inner batches since they might contain tuples that need to be
 989          * reassigned to later inner batches.
 990          *
 991          * 3. Similarly, if we have increased nbatch since starting the outer
 992          * scan, we have to rescan outer batches in case they contain tuples that
 993          * need to be reassigned.
 994          */
 995         curbatch++;
 996         while (curbatch < nbatch &&
 997                    (hashtable->outerBatchFile[curbatch] == NULL ||
 998                         hashtable->innerBatchFile[curbatch] == NULL))
 999         {
1000                 if (hashtable->outerBatchFile[curbatch] &&
1001                         HJ_FILL_OUTER(hjstate))
1002                         break;                          /* must process due to rule 1 */
1003                 if (hashtable->innerBatchFile[curbatch] &&
1004                         HJ_FILL_INNER(hjstate))
1005                         break;                          /* must process due to rule 1 */
1006                 if (hashtable->innerBatchFile[curbatch] &&
1007                         nbatch != hashtable->nbatch_original)
1008                         break;                          /* must process due to rule 2 */
1009                 if (hashtable->outerBatchFile[curbatch] &&
1010                         nbatch != hashtable->nbatch_outstart)
1011                         break;                          /* must process due to rule 3 */
1012                 /* We can ignore this batch. */
1013                 /* Release associated temp files right away. */
1014                 if (hashtable->innerBatchFile[curbatch])
1015                         BufFileClose(hashtable->innerBatchFile[curbatch]);
1016                 hashtable->innerBatchFile[curbatch] = NULL;
1017                 if (hashtable->outerBatchFile[curbatch])
1018                         BufFileClose(hashtable->outerBatchFile[curbatch]);
1019                 hashtable->outerBatchFile[curbatch] = NULL;
1020                 curbatch++;
1021         }
1022
1023         if (curbatch >= nbatch)
1024                 return false;                   /* no more batches */
1025
1026         hashtable->curbatch = curbatch;
1027
1028         /*
1029          * Reload the hash table with the new inner batch (which could be empty)
1030          */
1031         ExecHashTableReset(hashtable);
1032
1033         innerFile = hashtable->innerBatchFile[curbatch];
1034
1035         if (innerFile != NULL)
1036         {
1037                 if (BufFileSeek(innerFile, 0, 0L, SEEK_SET))
1038                         ereport(ERROR,
1039                                         (errcode_for_file_access(),
1040                                          errmsg("could not rewind hash-join temporary file: %m")));
1041
1042                 while ((slot = ExecHashJoinGetSavedTuple(hjstate,
1043                                                                                                  innerFile,
1044                                                                                                  &hashvalue,
1045                                                                                                  hjstate->hj_HashTupleSlot)))
1046                 {
1047                         /*
1048                          * NOTE: some tuples may be sent to future batches.  Also, it is
1049                          * possible for hashtable->nbatch to be increased here!
1050                          */
1051                         ExecHashTableInsert(hashtable, slot, hashvalue);
1052                 }
1053
1054                 /*
1055                  * after we build the hash table, the inner batch file is no longer
1056                  * needed
1057                  */
1058                 BufFileClose(innerFile);
1059                 hashtable->innerBatchFile[curbatch] = NULL;
1060         }
1061
1062         /*
1063          * Rewind outer batch file (if present), so that we can start reading it.
1064          */
1065         if (hashtable->outerBatchFile[curbatch] != NULL)
1066         {
1067                 if (BufFileSeek(hashtable->outerBatchFile[curbatch], 0, 0L, SEEK_SET))
1068                         ereport(ERROR,
1069                                         (errcode_for_file_access(),
1070                                          errmsg("could not rewind hash-join temporary file: %m")));
1071         }
1072
1073         return true;
1074 }
1075
1076 /*
1077  * Choose a batch to work on, and attach to it.  Returns true if successful,
1078  * false if there are no more batches.
1079  */
1080 static bool
1081 ExecParallelHashJoinNewBatch(HashJoinState *hjstate)
1082 {
1083         HashJoinTable hashtable = hjstate->hj_HashTable;
1084         int                     start_batchno;
1085         int                     batchno;
1086
1087         /*
1088          * If we started up so late that the batch tracking array has been freed
1089          * already by ExecHashTableDetach(), then we are finished.  See also
1090          * ExecParallelHashEnsureBatchAccessors().
1091          */
1092         if (hashtable->batches == NULL)
1093                 return false;
1094
1095         /*
1096          * If we were already attached to a batch, remember not to bother checking
1097          * it again, and detach from it (possibly freeing the hash table if we are
1098          * last to detach).
1099          */
1100         if (hashtable->curbatch >= 0)
1101         {
1102                 hashtable->batches[hashtable->curbatch].done = true;
1103                 ExecHashTableDetachBatch(hashtable);
1104         }
1105
1106         /*
1107          * Search for a batch that isn't done.  We use an atomic counter to start
1108          * our search at a different batch in every participant when there are
1109          * more batches than participants.
1110          */
1111         batchno = start_batchno =
1112                 pg_atomic_fetch_add_u32(&hashtable->parallel_state->distributor, 1) %
1113                 hashtable->nbatch;
1114         do
1115         {
1116                 uint32          hashvalue;
1117                 MinimalTuple tuple;
1118                 TupleTableSlot *slot;
1119
1120                 if (!hashtable->batches[batchno].done)
1121                 {
1122                         SharedTuplestoreAccessor *inner_tuples;
1123                         Barrier    *batch_barrier =
1124                         &hashtable->batches[batchno].shared->batch_barrier;
1125
1126                         switch (BarrierAttach(batch_barrier))
1127                         {
1128                                 case PHJ_BATCH_ELECTING:
1129
1130                                         /* One backend allocates the hash table. */
1131                                         if (BarrierArriveAndWait(batch_barrier,
1132                                                                                          WAIT_EVENT_HASH_BATCH_ELECTING))
1133                                                 ExecParallelHashTableAlloc(hashtable, batchno);
1134                                         /* Fall through. */
1135
1136                                 case PHJ_BATCH_ALLOCATING:
1137                                         /* Wait for allocation to complete. */
1138                                         BarrierArriveAndWait(batch_barrier,
1139                                                                                  WAIT_EVENT_HASH_BATCH_ALLOCATING);
1140                                         /* Fall through. */
1141
1142                                 case PHJ_BATCH_LOADING:
1143                                         /* Start (or join in) loading tuples. */
1144                                         ExecParallelHashTableSetCurrentBatch(hashtable, batchno);
1145                                         inner_tuples = hashtable->batches[batchno].inner_tuples;
1146                                         sts_begin_parallel_scan(inner_tuples);
1147                                         while ((tuple = sts_parallel_scan_next(inner_tuples,
1148                                                                                                                    &hashvalue)))
1149                                         {
1150                                                 slot = ExecStoreMinimalTuple(tuple,
1151                                                                                                          hjstate->hj_HashTupleSlot,
1152                                                                                                          false);
1153                                                 ExecParallelHashTableInsertCurrentBatch(hashtable, slot,
1154                                                                                                                                 hashvalue);
1155                                         }
1156                                         sts_end_parallel_scan(inner_tuples);
1157                                         BarrierArriveAndWait(batch_barrier,
1158                                                                                  WAIT_EVENT_HASH_BATCH_LOADING);
1159                                         /* Fall through. */
1160
1161                                 case PHJ_BATCH_PROBING:
1162
1163                                         /*
1164                                          * This batch is ready to probe.  Return control to
1165                                          * caller. We stay attached to batch_barrier so that the
1166                                          * hash table stays alive until everyone's finished
1167                                          * probing it, but no participant is allowed to wait at
1168                                          * this barrier again (or else a deadlock could occur).
1169                                          * All attached participants must eventually call
1170                                          * BarrierArriveAndDetach() so that the final phase
1171                                          * PHJ_BATCH_DONE can be reached.
1172                                          */
1173                                         ExecParallelHashTableSetCurrentBatch(hashtable, batchno);
1174                                         sts_begin_parallel_scan(hashtable->batches[batchno].outer_tuples);
1175                                         return true;
1176
1177                                 case PHJ_BATCH_DONE:
1178
1179                                         /*
1180                                          * Already done.  Detach and go around again (if any
1181                                          * remain).
1182                                          */
1183                                         BarrierDetach(batch_barrier);
1184                                         hashtable->batches[batchno].done = true;
1185                                         hashtable->curbatch = -1;
1186                                         break;
1187
1188                                 default:
1189                                         elog(ERROR, "unexpected batch phase %d",
1190                                                  BarrierPhase(batch_barrier));
1191                         }
1192                 }
1193                 batchno = (batchno + 1) % hashtable->nbatch;
1194         } while (batchno != start_batchno);
1195
1196         return false;
1197 }
1198
1199 /*
1200  * ExecHashJoinSaveTuple
1201  *              save a tuple to a batch file.
1202  *
1203  * The data recorded in the file for each tuple is its hash value,
1204  * then the tuple in MinimalTuple format.
1205  *
1206  * Note: it is important always to call this in the regular executor
1207  * context, not in a shorter-lived context; else the temp file buffers
1208  * will get messed up.
1209  */
1210 void
1211 ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
1212                                           BufFile **fileptr)
1213 {
1214         BufFile    *file = *fileptr;
1215         size_t          written;
1216
1217         if (file == NULL)
1218         {
1219                 /* First write to this batch file, so open it. */
1220                 file = BufFileCreateTemp(false);
1221                 *fileptr = file;
1222         }
1223
1224         written = BufFileWrite(file, (void *) &hashvalue, sizeof(uint32));
1225         if (written != sizeof(uint32))
1226                 ereport(ERROR,
1227                                 (errcode_for_file_access(),
1228                                  errmsg("could not write to hash-join temporary file: %m")));
1229
1230         written = BufFileWrite(file, (void *) tuple, tuple->t_len);
1231         if (written != tuple->t_len)
1232                 ereport(ERROR,
1233                                 (errcode_for_file_access(),
1234                                  errmsg("could not write to hash-join temporary file: %m")));
1235 }
1236
1237 /*
1238  * ExecHashJoinGetSavedTuple
1239  *              read the next tuple from a batch file.  Return NULL if no more.
1240  *
1241  * On success, *hashvalue is set to the tuple's hash value, and the tuple
1242  * itself is stored in the given slot.
1243  */
1244 static TupleTableSlot *
1245 ExecHashJoinGetSavedTuple(HashJoinState *hjstate,
1246                                                   BufFile *file,
1247                                                   uint32 *hashvalue,
1248                                                   TupleTableSlot *tupleSlot)
1249 {
1250         uint32          header[2];
1251         size_t          nread;
1252         MinimalTuple tuple;
1253
1254         /*
1255          * We check for interrupts here because this is typically taken as an
1256          * alternative code path to an ExecProcNode() call, which would include
1257          * such a check.
1258          */
1259         CHECK_FOR_INTERRUPTS();
1260
1261         /*
1262          * Since both the hash value and the MinimalTuple length word are uint32,
1263          * we can read them both in one BufFileRead() call without any type
1264          * cheating.
1265          */
1266         nread = BufFileRead(file, (void *) header, sizeof(header));
1267         if (nread == 0)                         /* end of file */
1268         {
1269                 ExecClearTuple(tupleSlot);
1270                 return NULL;
1271         }
1272         if (nread != sizeof(header))
1273                 ereport(ERROR,
1274                                 (errcode_for_file_access(),
1275                                  errmsg("could not read from hash-join temporary file: %m")));
1276         *hashvalue = header[0];
1277         tuple = (MinimalTuple) palloc(header[1]);
1278         tuple->t_len = header[1];
1279         nread = BufFileRead(file,
1280                                                 (void *) ((char *) tuple + sizeof(uint32)),
1281                                                 header[1] - sizeof(uint32));
1282         if (nread != header[1] - sizeof(uint32))
1283                 ereport(ERROR,
1284                                 (errcode_for_file_access(),
1285                                  errmsg("could not read from hash-join temporary file: %m")));
1286         return ExecStoreMinimalTuple(tuple, tupleSlot, true);
1287 }
1288
1289
1290 void
1291 ExecReScanHashJoin(HashJoinState *node)
1292 {
1293         /*
1294          * In a multi-batch join, we currently have to do rescans the hard way,
1295          * primarily because batch temp files may have already been released. But
1296          * if it's a single-batch join, and there is no parameter change for the
1297          * inner subnode, then we can just re-use the existing hash table without
1298          * rebuilding it.
1299          */
1300         if (node->hj_HashTable != NULL)
1301         {
1302                 if (node->hj_HashTable->nbatch == 1 &&
1303                         node->js.ps.righttree->chgParam == NULL)
1304                 {
1305                         /*
1306                          * Okay to reuse the hash table; needn't rescan inner, either.
1307                          *
1308                          * However, if it's a right/full join, we'd better reset the
1309                          * inner-tuple match flags contained in the table.
1310                          */
1311                         if (HJ_FILL_INNER(node))
1312                                 ExecHashTableResetMatchFlags(node->hj_HashTable);
1313
1314                         /*
1315                          * Also, we need to reset our state about the emptiness of the
1316                          * outer relation, so that the new scan of the outer will update
1317                          * it correctly if it turns out to be empty this time. (There's no
1318                          * harm in clearing it now because ExecHashJoin won't need the
1319                          * info.  In the other cases, where the hash table doesn't exist
1320                          * or we are destroying it, we leave this state alone because
1321                          * ExecHashJoin will need it the first time through.)
1322                          */
1323                         node->hj_OuterNotEmpty = false;
1324
1325                         /* ExecHashJoin can skip the BUILD_HASHTABLE step */
1326                         node->hj_JoinState = HJ_NEED_NEW_OUTER;
1327                 }
1328                 else
1329                 {
1330                         /* must destroy and rebuild hash table */
1331                         ExecHashTableDestroy(node->hj_HashTable);
1332                         node->hj_HashTable = NULL;
1333                         node->hj_JoinState = HJ_BUILD_HASHTABLE;
1334
1335                         /*
1336                          * if chgParam of subnode is not null then plan will be re-scanned
1337                          * by first ExecProcNode.
1338                          */
1339                         if (node->js.ps.righttree->chgParam == NULL)
1340                                 ExecReScan(node->js.ps.righttree);
1341                 }
1342         }
1343
1344         /* Always reset intra-tuple state */
1345         node->hj_CurHashValue = 0;
1346         node->hj_CurBucketNo = 0;
1347         node->hj_CurSkewBucketNo = INVALID_SKEW_BUCKET_NO;
1348         node->hj_CurTuple = NULL;
1349
1350         node->hj_MatchedOuter = false;
1351         node->hj_FirstOuterTupleSlot = NULL;
1352
1353         /*
1354          * if chgParam of subnode is not null then plan will be re-scanned by
1355          * first ExecProcNode.
1356          */
1357         if (node->js.ps.lefttree->chgParam == NULL)
1358                 ExecReScan(node->js.ps.lefttree);
1359 }
1360
1361 void
1362 ExecShutdownHashJoin(HashJoinState *node)
1363 {
1364         if (node->hj_HashTable)
1365         {
1366                 /*
1367                  * Detach from shared state before DSM memory goes away.  This makes
1368                  * sure that we don't have any pointers into DSM memory by the time
1369                  * ExecEndHashJoin runs.
1370                  */
1371                 ExecHashTableDetachBatch(node->hj_HashTable);
1372                 ExecHashTableDetach(node->hj_HashTable);
1373         }
1374 }
1375
1376 static void
1377 ExecParallelHashJoinPartitionOuter(HashJoinState *hjstate)
1378 {
1379         PlanState  *outerState = outerPlanState(hjstate);
1380         ExprContext *econtext = hjstate->js.ps.ps_ExprContext;
1381         HashJoinTable hashtable = hjstate->hj_HashTable;
1382         TupleTableSlot *slot;
1383         uint32          hashvalue;
1384         int                     i;
1385
1386         Assert(hjstate->hj_FirstOuterTupleSlot == NULL);
1387
1388         /* Execute outer plan, writing all tuples to shared tuplestores. */
1389         for (;;)
1390         {
1391                 slot = ExecProcNode(outerState);
1392                 if (TupIsNull(slot))
1393                         break;
1394                 econtext->ecxt_outertuple = slot;
1395                 if (ExecHashGetHashValue(hashtable, econtext,
1396                                                                  hjstate->hj_OuterHashKeys,
1397                                                                  true,  /* outer tuple */
1398                                                                  false, /* outer join, currently unsupported */
1399                                                                  &hashvalue))
1400                 {
1401                         int                     batchno;
1402                         int                     bucketno;
1403
1404                         ExecHashGetBucketAndBatch(hashtable, hashvalue, &bucketno,
1405                                                                           &batchno);
1406                         sts_puttuple(hashtable->batches[batchno].outer_tuples,
1407                                                  &hashvalue, ExecFetchSlotMinimalTuple(slot));
1408                 }
1409                 CHECK_FOR_INTERRUPTS();
1410         }
1411
1412         /* Make sure all outer partitions are readable by any backend. */
1413         for (i = 0; i < hashtable->nbatch; ++i)
1414                 sts_end_write(hashtable->batches[i].outer_tuples);
1415 }
1416
1417 void
1418 ExecHashJoinEstimate(HashJoinState *state, ParallelContext *pcxt)
1419 {
1420         shm_toc_estimate_chunk(&pcxt->estimator, sizeof(ParallelHashJoinState));
1421         shm_toc_estimate_keys(&pcxt->estimator, 1);
1422 }
1423
1424 void
1425 ExecHashJoinInitializeDSM(HashJoinState *state, ParallelContext *pcxt)
1426 {
1427         int                     plan_node_id = state->js.ps.plan->plan_node_id;
1428         HashState  *hashNode;
1429         ParallelHashJoinState *pstate;
1430
1431         /*
1432          * Disable shared hash table mode if we failed to create a real DSM
1433          * segment, because that means that we don't have a DSA area to work with.
1434          */
1435         if (pcxt->seg == NULL)
1436                 return;
1437
1438         ExecSetExecProcNode(&state->js.ps, ExecParallelHashJoin);
1439
1440         /*
1441          * Set up the state needed to coordinate access to the shared hash
1442          * table(s), using the plan node ID as the toc key.
1443          */
1444         pstate = shm_toc_allocate(pcxt->toc, sizeof(ParallelHashJoinState));
1445         shm_toc_insert(pcxt->toc, plan_node_id, pstate);
1446
1447         /*
1448          * Set up the shared hash join state with no batches initially.
1449          * ExecHashTableCreate() will prepare at least one later and set nbatch
1450          * and space_allowed.
1451          */
1452         pstate->nbatch = 0;
1453         pstate->space_allowed = 0;
1454         pstate->batches = InvalidDsaPointer;
1455         pstate->old_batches = InvalidDsaPointer;
1456         pstate->nbuckets = 0;
1457         pstate->growth = PHJ_GROWTH_OK;
1458         pstate->chunk_work_queue = InvalidDsaPointer;
1459         pg_atomic_init_u32(&pstate->distributor, 0);
1460         pstate->nparticipants = pcxt->nworkers + 1;
1461         pstate->total_tuples = 0;
1462         LWLockInitialize(&pstate->lock,
1463                                          LWTRANCHE_PARALLEL_HASH_JOIN);
1464         BarrierInit(&pstate->build_barrier, 0);
1465         BarrierInit(&pstate->grow_batches_barrier, 0);
1466         BarrierInit(&pstate->grow_buckets_barrier, 0);
1467
1468         /* Set up the space we'll use for shared temporary files. */
1469         SharedFileSetInit(&pstate->fileset, pcxt->seg);
1470
1471         /* Initialize the shared state in the hash node. */
1472         hashNode = (HashState *) innerPlanState(state);
1473         hashNode->parallel_state = pstate;
1474 }
1475
1476 /* ----------------------------------------------------------------
1477  *              ExecHashJoinReInitializeDSM
1478  *
1479  *              Reset shared state before beginning a fresh scan.
1480  * ----------------------------------------------------------------
1481  */
1482 void
1483 ExecHashJoinReInitializeDSM(HashJoinState *state, ParallelContext *cxt)
1484 {
1485         int                     plan_node_id = state->js.ps.plan->plan_node_id;
1486         ParallelHashJoinState *pstate =
1487         shm_toc_lookup(cxt->toc, plan_node_id, false);
1488
1489         /*
1490          * It would be possible to reuse the shared hash table in single-batch
1491          * cases by resetting and then fast-forwarding build_barrier to
1492          * PHJ_BUILD_DONE and batch 0's batch_barrier to PHJ_BATCH_PROBING, but
1493          * currently shared hash tables are already freed by now (by the last
1494          * participant to detach from the batch).  We could consider keeping it
1495          * around for single-batch joins.  We'd also need to adjust
1496          * finalize_plan() so that it doesn't record a dummy dependency for
1497          * Parallel Hash nodes, preventing the rescan optimization.  For now we
1498          * don't try.
1499          */
1500
1501         /* Detach, freeing any remaining shared memory. */
1502         if (state->hj_HashTable != NULL)
1503         {
1504                 ExecHashTableDetachBatch(state->hj_HashTable);
1505                 ExecHashTableDetach(state->hj_HashTable);
1506         }
1507
1508         /* Clear any shared batch files. */
1509         SharedFileSetDeleteAll(&pstate->fileset);
1510
1511         /* Reset build_barrier to PHJ_BUILD_ELECTING so we can go around again. */
1512         BarrierInit(&pstate->build_barrier, 0);
1513 }
1514
1515 void
1516 ExecHashJoinInitializeWorker(HashJoinState *state,
1517                                                          ParallelWorkerContext *pwcxt)
1518 {
1519         HashState  *hashNode;
1520         int                     plan_node_id = state->js.ps.plan->plan_node_id;
1521         ParallelHashJoinState *pstate =
1522         shm_toc_lookup(pwcxt->toc, plan_node_id, false);
1523
1524         /* Attach to the space for shared temporary files. */
1525         SharedFileSetAttach(&pstate->fileset, pwcxt->seg);
1526
1527         /* Attach to the shared state in the hash node. */
1528         hashNode = (HashState *) innerPlanState(state);
1529         hashNode->parallel_state = pstate;
1530
1531         ExecSetExecProcNode(&state->js.ps, ExecParallelHashJoin);
1532 }