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

   1 /*-------------------------------------------------------------------------
   2  *
   3  * execMain.c
   4  *        top level executor interface routines
   5  *
   6  * INTERFACE ROUTINES
   7  *      ExecutorStart()
   8  *      ExecutorRun()
   9  *      ExecutorEnd()
  10  *
  11  *      The old ExecutorMain() has been replaced by ExecutorStart(),
  12  *      ExecutorRun() and ExecutorEnd()
  13  *
  14  *      These three procedures are the external interfaces to the executor.
  15  *      In each case, the query descriptor is required as an argument.
  16  *
  17  *      ExecutorStart() must be called at the beginning of execution of any
  18  *      query plan and ExecutorEnd() should always be called at the end of
  19  *      execution of a plan.
  20  *
  21  *      ExecutorRun accepts direction and count arguments that specify whether
  22  *      the plan is to be executed forwards, backwards, and for how many tuples.
  23  *
  24  * Portions Copyright (c) 1996-2003, PostgreSQL Global Development Group
  25  * Portions Copyright (c) 1994, Regents of the University of California
  26  *
  27  *
  28  * IDENTIFICATION
  29  *        $PostgreSQL: pgsql/src/backend/executor/execMain.c,v 1.229 2004/03/02 18:56:15 tgl Exp $
  30  *
  31  *-------------------------------------------------------------------------
  32  */
  33 #include "postgres.h"
  34
  35 #include "access/heapam.h"
  36 #include "catalog/heap.h"
  37 #include "catalog/namespace.h"
  38 #include "commands/tablecmds.h"
  39 #include "commands/trigger.h"
  40 #include "executor/execdebug.h"
  41 #include "executor/execdefs.h"
  42 #include "miscadmin.h"
  43 #include "optimizer/clauses.h"
  44 #include "optimizer/var.h"
  45 #include "parser/parsetree.h"
  46 #include "utils/acl.h"
  47 #include "utils/guc.h"
  48 #include "utils/lsyscache.h"
  49
  50
  51 typedef struct execRowMark
  52 {
  53         Relation        relation;
  54         Index           rti;
  55         char            resname[32];
  56 } execRowMark;
  57
  58 typedef struct evalPlanQual
  59 {
  60         Index           rti;
  61         EState     *estate;
  62         PlanState  *planstate;
  63         struct evalPlanQual *next;      /* stack of active PlanQual plans */
  64         struct evalPlanQual *free;      /* list of free PlanQual plans */
  65 } evalPlanQual;
  66
  67 /* decls for local routines only used within this module */
  68 static void InitPlan(QueryDesc *queryDesc, bool explainOnly);
  69 static void initResultRelInfo(ResultRelInfo *resultRelInfo,
  70                                   Index resultRelationIndex,
  71                                   List *rangeTable,
  72                                   CmdType operation);
  73 static TupleTableSlot *ExecutePlan(EState *estate, PlanState *planstate,
  74                         CmdType operation,
  75                         long numberTuples,
  76                         ScanDirection direction,
  77                         DestReceiver *dest);
  78 static void ExecSelect(TupleTableSlot *slot,
  79                    DestReceiver *dest,
  80                    EState *estate);
  81 static void ExecInsert(TupleTableSlot *slot, ItemPointer tupleid,
  82                    EState *estate);
  83 static void ExecDelete(TupleTableSlot *slot, ItemPointer tupleid,
  84                    EState *estate);
  85 static void ExecUpdate(TupleTableSlot *slot, ItemPointer tupleid,
  86                    EState *estate);
  87 static TupleTableSlot *EvalPlanQualNext(EState *estate);
  88 static void EndEvalPlanQual(EState *estate);
  89 static void ExecCheckRTEPerms(RangeTblEntry *rte);
  90 static void ExecCheckXactReadOnly(Query *parsetree);
  91 static void EvalPlanQualStart(evalPlanQual *epq, EState *estate,
  92                                   evalPlanQual *priorepq);
  93 static void EvalPlanQualStop(evalPlanQual *epq);
  94
  95 /* end of local decls */
  96
  97
  98 /* ----------------------------------------------------------------
  99  *              ExecutorStart
 100  *
 101  *              This routine must be called at the beginning of any execution of any
 102  *              query plan
 103  *
 104  * Takes a QueryDesc previously created by CreateQueryDesc (it's not real
 105  * clear why we bother to separate the two functions, but...).  The tupDesc
 106  * field of the QueryDesc is filled in to describe the tuples that will be
 107  * returned, and the internal fields (estate and planstate) are set up.
 108  *
 109  * If useCurrentSnapshot is true, run the query with the latest available
 110  * snapshot, instead of the normal QuerySnapshot.  Also, if it's an update
 111  * or delete query, check that the rows to be updated or deleted would be
 112  * visible to the normal QuerySnapshot.  (This is a special-case behavior
 113  * needed for referential integrity updates in serializable transactions.
 114  * We must check all currently-committed rows, but we want to throw a
 115  * can't-serialize error if any rows that would need updates would not be
 116  * visible under the normal serializable snapshot.)
 117  *
 118  * If explainOnly is true, we are not actually intending to run the plan,
 119  * only to set up for EXPLAIN; so skip unwanted side-effects.
 120  *
 121  * NB: the CurrentMemoryContext when this is called will become the parent
 122  * of the per-query context used for this Executor invocation.
 123  * ----------------------------------------------------------------
 124  */
 125 void
 126 ExecutorStart(QueryDesc *queryDesc, bool useCurrentSnapshot, bool explainOnly)
 127 {
 128         EState     *estate;
 129         MemoryContext oldcontext;
 130
 131         /* sanity checks: queryDesc must not be started already */
 132         Assert(queryDesc != NULL);
 133         Assert(queryDesc->estate == NULL);
 134
 135         /*
 136          * If the transaction is read-only, we need to check if any writes are
 137          * planned to non-temporary tables.
 138          */
 139         if (XactReadOnly && !explainOnly)
 140                 ExecCheckXactReadOnly(queryDesc->parsetree);
 141
 142         /*
 143          * Build EState, switch into per-query memory context for startup.
 144          */
 145         estate = CreateExecutorState();
 146         queryDesc->estate = estate;
 147
 148         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 149
 150         /*
 151          * Fill in parameters, if any, from queryDesc
 152          */
 153         estate->es_param_list_info = queryDesc->params;
 154
 155         if (queryDesc->plantree->nParamExec > 0)
 156                 estate->es_param_exec_vals = (ParamExecData *)
 157                         palloc0(queryDesc->plantree->nParamExec * sizeof(ParamExecData));
 158
 159         estate->es_instrument = queryDesc->doInstrument;
 160
 161         /*
 162          * Make our own private copy of the current query snapshot data.
 163          *
 164          * This "freezes" our idea of which tuples are good and which are not for
 165          * the life of this query, even if it outlives the current command and
 166          * current snapshot.
 167          */
 168         if (useCurrentSnapshot)
 169         {
 170                 /* RI update/delete query --- must use an up-to-date snapshot */
 171                 estate->es_snapshot = CopyCurrentSnapshot();
 172                 /* crosscheck updates/deletes against transaction snapshot */
 173                 estate->es_crosscheck_snapshot = CopyQuerySnapshot();
 174         }
 175         else
 176         {
 177                 /* normal query --- use query snapshot, no crosscheck */
 178                 estate->es_snapshot = CopyQuerySnapshot();
 179                 estate->es_crosscheck_snapshot = SnapshotAny;
 180         }
 181
 182         /*
 183          * Initialize the plan state tree
 184          */
 185         InitPlan(queryDesc, explainOnly);
 186
 187         MemoryContextSwitchTo(oldcontext);
 188 }
 189
 190 /* ----------------------------------------------------------------
 191  *              ExecutorRun
 192  *
 193  *              This is the main routine of the executor module. It accepts
 194  *              the query descriptor from the traffic cop and executes the
 195  *              query plan.
 196  *
 197  *              ExecutorStart must have been called already.
 198  *
 199  *              If direction is NoMovementScanDirection then nothing is done
 200  *              except to start up/shut down the destination.  Otherwise,
 201  *              we retrieve up to 'count' tuples in the specified direction.
 202  *
 203  *              Note: count = 0 is interpreted as no portal limit, i.e., run to
 204  *              completion.
 205  *
 206  * ----------------------------------------------------------------
 207  */
 208 TupleTableSlot *
 209 ExecutorRun(QueryDesc *queryDesc,
 210                         ScanDirection direction, long count)
 211 {
 212         EState     *estate;
 213         CmdType         operation;
 214         DestReceiver *dest;
 215         TupleTableSlot *result;
 216         MemoryContext oldcontext;
 217
 218         /* sanity checks */
 219         Assert(queryDesc != NULL);
 220
 221         estate = queryDesc->estate;
 222
 223         Assert(estate != NULL);
 224
 225         /*
 226          * Switch into per-query memory context
 227          */
 228         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 229
 230         /*
 231          * extract information from the query descriptor and the query
 232          * feature.
 233          */
 234         operation = queryDesc->operation;
 235         dest = queryDesc->dest;
 236
 237         /*
 238          * startup tuple receiver
 239          */
 240         estate->es_processed = 0;
 241         estate->es_lastoid = InvalidOid;
 242
 243         (*dest->rStartup) (dest, operation, queryDesc->tupDesc);
 244
 245         /*
 246          * run plan
 247          */
 248         if (direction == NoMovementScanDirection)
 249                 result = NULL;
 250         else
 251                 result = ExecutePlan(estate,
 252                                                          queryDesc->planstate,
 253                                                          operation,
 254                                                          count,
 255                                                          direction,
 256                                                          dest);
 257
 258         /*
 259          * shutdown receiver
 260          */
 261         (*dest->rShutdown) (dest);
 262
 263         MemoryContextSwitchTo(oldcontext);
 264
 265         return result;
 266 }
 267
 268 /* ----------------------------------------------------------------
 269  *              ExecutorEnd
 270  *
 271  *              This routine must be called at the end of execution of any
 272  *              query plan
 273  * ----------------------------------------------------------------
 274  */
 275 void
 276 ExecutorEnd(QueryDesc *queryDesc)
 277 {
 278         EState     *estate;
 279         MemoryContext oldcontext;
 280
 281         /* sanity checks */
 282         Assert(queryDesc != NULL);
 283
 284         estate = queryDesc->estate;
 285
 286         Assert(estate != NULL);
 287
 288         /*
 289          * Switch into per-query memory context to run ExecEndPlan
 290          */
 291         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 292
 293         ExecEndPlan(queryDesc->planstate, estate);
 294
 295         /*
 296          * Must switch out of context before destroying it
 297          */
 298         MemoryContextSwitchTo(oldcontext);
 299
 300         /*
 301          * Release EState and per-query memory context.  This should release
 302          * everything the executor has allocated.
 303          */
 304         FreeExecutorState(estate);
 305
 306         /* Reset queryDesc fields that no longer point to anything */
 307         queryDesc->tupDesc = NULL;
 308         queryDesc->estate = NULL;
 309         queryDesc->planstate = NULL;
 310 }
 311
 312 /* ----------------------------------------------------------------
 313  *              ExecutorRewind
 314  *
 315  *              This routine may be called on an open queryDesc to rewind it
 316  *              to the start.
 317  * ----------------------------------------------------------------
 318  */
 319 void
 320 ExecutorRewind(QueryDesc *queryDesc)
 321 {
 322         EState     *estate;
 323         MemoryContext oldcontext;
 324
 325         /* sanity checks */
 326         Assert(queryDesc != NULL);
 327
 328         estate = queryDesc->estate;
 329
 330         Assert(estate != NULL);
 331
 332         /* It's probably not sensible to rescan updating queries */
 333         Assert(queryDesc->operation == CMD_SELECT);
 334
 335         /*
 336          * Switch into per-query memory context
 337          */
 338         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 339
 340         /*
 341          * rescan plan
 342          */
 343         ExecReScan(queryDesc->planstate, NULL);
 344
 345         MemoryContextSwitchTo(oldcontext);
 346 }
 347
 348
 349 /*
 350  * ExecCheckRTPerms
 351  *              Check access permissions for all relations listed in a range table.
 352  */
 353 void
 354 ExecCheckRTPerms(List *rangeTable)
 355 {
 356         List       *lp;
 357
 358         foreach(lp, rangeTable)
 359         {
 360                 RangeTblEntry *rte = lfirst(lp);
 361
 362                 ExecCheckRTEPerms(rte);
 363         }
 364 }
 365
 366 /*
 367  * ExecCheckRTEPerms
 368  *              Check access permissions for a single RTE.
 369  */
 370 static void
 371 ExecCheckRTEPerms(RangeTblEntry *rte)
 372 {
 373         AclMode         requiredPerms;
 374         Oid                     relOid;
 375         AclId           userid;
 376
 377         /*
 378          * If it's a subquery, recursively examine its rangetable.
 379          */
 380         if (rte->rtekind == RTE_SUBQUERY)
 381         {
 382                 ExecCheckRTPerms(rte->subquery->rtable);
 383                 return;
 384         }
 385
 386         /*
 387          * Otherwise, only plain-relation RTEs need to be checked here.
 388          * Function RTEs are checked by init_fcache when the function is
 389          * prepared for execution. Join and special RTEs need no checks.
 390          */
 391         if (rte->rtekind != RTE_RELATION)
 392                 return;
 393
 394         /*
 395          * No work if requiredPerms is empty.
 396          */
 397         requiredPerms = rte->requiredPerms;
 398         if (requiredPerms == 0)
 399                 return;
 400
 401         relOid = rte->relid;
 402
 403         /*
 404          * userid to check as: current user unless we have a setuid
 405          * indication.
 406          *
 407          * Note: GetUserId() is presently fast enough that there's no harm in
 408          * calling it separately for each RTE.  If that stops being true, we
 409          * could call it once in ExecCheckRTPerms and pass the userid down
 410          * from there.  But for now, no need for the extra clutter.
 411          */
 412         userid = rte->checkAsUser ? rte->checkAsUser : GetUserId();
 413
 414         /*
 415          * For each bit in requiredPerms, apply the required check.  (We can't
 416          * do this in one aclcheck call because aclcheck treats multiple bits
 417          * as OR semantics, when we want AND.)
 418          *
 419          * We use a well-known cute trick for isolating the rightmost one-bit
 420          * in a nonzero word.  See nodes/bitmapset.c for commentary.
 421          */
 422 #define RIGHTMOST_ONE(x) ((int32) (x) & -((int32) (x)))
 423
 424         while (requiredPerms != 0)
 425         {
 426                 AclMode         thisPerm;
 427                 AclResult       aclcheck_result;
 428
 429                 thisPerm = RIGHTMOST_ONE(requiredPerms);
 430                 requiredPerms &= ~thisPerm;
 431
 432                 aclcheck_result = pg_class_aclcheck(relOid, userid, thisPerm);
 433                 if (aclcheck_result != ACLCHECK_OK)
 434                         aclcheck_error(aclcheck_result, ACL_KIND_CLASS,
 435                                                    get_rel_name(relOid));
 436         }
 437 }
 438
 439 /*
 440  * Check that the query does not imply any writes to non-temp tables.
 441  */
 442 static void
 443 ExecCheckXactReadOnly(Query *parsetree)
 444 {
 445         List       *lp;
 446
 447         /*
 448          * CREATE TABLE AS or SELECT INTO?
 449          *
 450          * XXX should we allow this if the destination is temp?
 451          */
 452         if (parsetree->into != NULL)
 453                 goto fail;
 454
 455         /* Fail if write permissions are requested on any non-temp table */
 456         foreach(lp, parsetree->rtable)
 457         {
 458                 RangeTblEntry *rte = lfirst(lp);
 459
 460                 if (rte->rtekind == RTE_SUBQUERY)
 461                 {
 462                         ExecCheckXactReadOnly(rte->subquery);
 463                         continue;
 464                 }
 465
 466                 if (rte->rtekind != RTE_RELATION)
 467                         continue;
 468
 469                 if ((rte->requiredPerms & (~ACL_SELECT)) == 0)
 470                         continue;
 471
 472                 if (isTempNamespace(get_rel_namespace(rte->relid)))
 473                         continue;
 474
 475                 goto fail;
 476         }
 477
 478         return;
 479
 480 fail:
 481         ereport(ERROR,
 482                         (errcode(ERRCODE_READ_ONLY_SQL_TRANSACTION),
 483                          errmsg("transaction is read-only")));
 484 }
 485
 486
 487 /* ----------------------------------------------------------------
 488  *              InitPlan
 489  *
 490  *              Initializes the query plan: open files, allocate storage
 491  *              and start up the rule manager
 492  * ----------------------------------------------------------------
 493  */
 494 static void
 495 InitPlan(QueryDesc *queryDesc, bool explainOnly)
 496 {
 497         CmdType         operation = queryDesc->operation;
 498         Query      *parseTree = queryDesc->parsetree;
 499         Plan       *plan = queryDesc->plantree;
 500         EState     *estate = queryDesc->estate;
 501         PlanState  *planstate;
 502         List       *rangeTable;
 503         Relation        intoRelationDesc;
 504         bool            do_select_into;
 505         TupleDesc       tupType;
 506
 507         /*
 508          * Do permissions checks.  It's sufficient to examine the query's top
 509          * rangetable here --- subplan RTEs will be checked during
 510          * ExecInitSubPlan().
 511          */
 512         ExecCheckRTPerms(parseTree->rtable);
 513
 514         /*
 515          * get information from query descriptor
 516          */
 517         rangeTable = parseTree->rtable;
 518
 519         /*
 520          * initialize the node's execution state
 521          */
 522         estate->es_range_table = rangeTable;
 523
 524         /*
 525          * if there is a result relation, initialize result relation stuff
 526          */
 527         if (parseTree->resultRelation != 0 && operation != CMD_SELECT)
 528         {
 529                 List       *resultRelations = parseTree->resultRelations;
 530                 int                     numResultRelations;
 531                 ResultRelInfo *resultRelInfos;
 532
 533                 if (resultRelations != NIL)
 534                 {
 535                         /*
 536                          * Multiple result relations (due to inheritance)
 537                          * parseTree->resultRelations identifies them all
 538                          */
 539                         ResultRelInfo *resultRelInfo;
 540
 541                         numResultRelations = length(resultRelations);
 542                         resultRelInfos = (ResultRelInfo *)
 543                                 palloc(numResultRelations * sizeof(ResultRelInfo));
 544                         resultRelInfo = resultRelInfos;
 545                         while (resultRelations != NIL)
 546                         {
 547                                 initResultRelInfo(resultRelInfo,
 548                                                                   lfirsti(resultRelations),
 549                                                                   rangeTable,
 550                                                                   operation);
 551                                 resultRelInfo++;
 552                                 resultRelations = lnext(resultRelations);
 553                         }
 554                 }
 555                 else
 556                 {
 557                         /*
 558                          * Single result relation identified by
 559                          * parseTree->resultRelation
 560                          */
 561                         numResultRelations = 1;
 562                         resultRelInfos = (ResultRelInfo *) palloc(sizeof(ResultRelInfo));
 563                         initResultRelInfo(resultRelInfos,
 564                                                           parseTree->resultRelation,
 565                                                           rangeTable,
 566                                                           operation);
 567                 }
 568
 569                 estate->es_result_relations = resultRelInfos;
 570                 estate->es_num_result_relations = numResultRelations;
 571                 /* Initialize to first or only result rel */
 572                 estate->es_result_relation_info = resultRelInfos;
 573         }
 574         else
 575         {
 576                 /*
 577                  * if no result relation, then set state appropriately
 578                  */
 579                 estate->es_result_relations = NULL;
 580                 estate->es_num_result_relations = 0;
 581                 estate->es_result_relation_info = NULL;
 582         }
 583
 584         /*
 585          * Detect whether we're doing SELECT INTO.  If so, set the force_oids
 586          * flag appropriately so that the plan tree will be initialized with
 587          * the correct tuple descriptors.
 588          */
 589         do_select_into = false;
 590
 591         if (operation == CMD_SELECT && parseTree->into != NULL)
 592         {
 593                 do_select_into = true;
 594                 estate->es_select_into = true;
 595                 estate->es_into_oids = parseTree->intoHasOids;
 596         }
 597
 598         /*
 599          * Have to lock relations selected for update
 600          */
 601         estate->es_rowMark = NIL;
 602         if (parseTree->rowMarks != NIL)
 603         {
 604                 List       *l;
 605
 606                 foreach(l, parseTree->rowMarks)
 607                 {
 608                         Index           rti = lfirsti(l);
 609                         Oid                     relid = getrelid(rti, rangeTable);
 610                         Relation        relation;
 611                         execRowMark *erm;
 612
 613                         relation = heap_open(relid, RowShareLock);
 614                         erm = (execRowMark *) palloc(sizeof(execRowMark));
 615                         erm->relation = relation;
 616                         erm->rti = rti;
 617                         snprintf(erm->resname, sizeof(erm->resname), "ctid%u", rti);
 618                         estate->es_rowMark = lappend(estate->es_rowMark, erm);
 619                 }
 620         }
 621
 622         /*
 623          * initialize the executor "tuple" table.  We need slots for all the
 624          * plan nodes, plus possibly output slots for the junkfilter(s). At
 625          * this point we aren't sure if we need junkfilters, so just add slots
 626          * for them unconditionally.
 627          */
 628         {
 629                 int                     nSlots = ExecCountSlotsNode(plan);
 630
 631                 if (parseTree->resultRelations != NIL)
 632                         nSlots += length(parseTree->resultRelations);
 633                 else
 634                         nSlots += 1;
 635                 estate->es_tupleTable = ExecCreateTupleTable(nSlots);
 636         }
 637
 638         /* mark EvalPlanQual not active */
 639         estate->es_topPlan = plan;
 640         estate->es_evalPlanQual = NULL;
 641         estate->es_evTupleNull = NULL;
 642         estate->es_evTuple = NULL;
 643         estate->es_useEvalPlan = false;
 644
 645         /*
 646          * initialize the private state information for all the nodes in the
 647          * query tree.  This opens files, allocates storage and leaves us
 648          * ready to start processing tuples.
 649          */
 650         planstate = ExecInitNode(plan, estate);
 651
 652         /*
 653          * Get the tuple descriptor describing the type of tuples to return.
 654          * (this is especially important if we are creating a relation with
 655          * "SELECT INTO")
 656          */
 657         tupType = ExecGetResultType(planstate);
 658
 659         /*
 660          * Initialize the junk filter if needed.  SELECT and INSERT queries
 661          * need a filter if there are any junk attrs in the tlist.      INSERT and
 662          * SELECT INTO also need a filter if the plan may return raw disk tuples
 663          * (else heap_insert will be scribbling on the source relation!).
 664          * UPDATE and DELETE always need a filter, since there's always a junk
 665          * 'ctid' attribute present --- no need to look first.
 666          */
 667         {
 668                 bool            junk_filter_needed = false;
 669                 List       *tlist;
 670
 671                 switch (operation)
 672                 {
 673                         case CMD_SELECT:
 674                         case CMD_INSERT:
 675                                 foreach(tlist, plan->targetlist)
 676                                 {
 677                                         TargetEntry *tle = (TargetEntry *) lfirst(tlist);
 678
 679                                         if (tle->resdom->resjunk)
 680                                         {
 681                                                 junk_filter_needed = true;
 682                                                 break;
 683                                         }
 684                                 }
 685                                 if (!junk_filter_needed &&
 686                                         (operation == CMD_INSERT || do_select_into) &&
 687                                         ExecMayReturnRawTuples(planstate))
 688                                         junk_filter_needed = true;
 689                                 break;
 690                         case CMD_UPDATE:
 691                         case CMD_DELETE:
 692                                 junk_filter_needed = true;
 693                                 break;
 694                         default:
 695                                 break;
 696                 }
 697
 698                 if (junk_filter_needed)
 699                 {
 700                         /*
 701                          * If there are multiple result relations, each one needs its
 702                          * own junk filter.  Note this is only possible for
 703                          * UPDATE/DELETE, so we can't be fooled by some needing a
 704                          * filter and some not.
 705                          */
 706                         if (parseTree->resultRelations != NIL)
 707                         {
 708                                 PlanState **appendplans;
 709                                 int                     as_nplans;
 710                                 ResultRelInfo *resultRelInfo;
 711                                 int                     i;
 712
 713                                 /* Top plan had better be an Append here. */
 714                                 Assert(IsA(plan, Append));
 715                                 Assert(((Append *) plan)->isTarget);
 716                                 Assert(IsA(planstate, AppendState));
 717                                 appendplans = ((AppendState *) planstate)->appendplans;
 718                                 as_nplans = ((AppendState *) planstate)->as_nplans;
 719                                 Assert(as_nplans == estate->es_num_result_relations);
 720                                 resultRelInfo = estate->es_result_relations;
 721                                 for (i = 0; i < as_nplans; i++)
 722                                 {
 723                                         PlanState  *subplan = appendplans[i];
 724                                         JunkFilter *j;
 725
 726                                         j = ExecInitJunkFilter(subplan->plan->targetlist,
 727                                                                                    ExecGetResultType(subplan),
 728                                                           ExecAllocTableSlot(estate->es_tupleTable));
 729                                         resultRelInfo->ri_junkFilter = j;
 730                                         resultRelInfo++;
 731                                 }
 732
 733                                 /*
 734                                  * Set active junkfilter too; at this point ExecInitAppend
 735                                  * has already selected an active result relation...
 736                                  */
 737                                 estate->es_junkFilter =
 738                                         estate->es_result_relation_info->ri_junkFilter;
 739                         }
 740                         else
 741                         {
 742                                 /* Normal case with just one JunkFilter */
 743                                 JunkFilter *j;
 744
 745                                 j = ExecInitJunkFilter(planstate->plan->targetlist,
 746                                                                            tupType,
 747                                                           ExecAllocTableSlot(estate->es_tupleTable));
 748                                 estate->es_junkFilter = j;
 749                                 if (estate->es_result_relation_info)
 750                                         estate->es_result_relation_info->ri_junkFilter = j;
 751
 752                                 /* For SELECT, want to return the cleaned tuple type */
 753                                 if (operation == CMD_SELECT)
 754                                         tupType = j->jf_cleanTupType;
 755                         }
 756                 }
 757                 else
 758                         estate->es_junkFilter = NULL;
 759         }
 760
 761         /*
 762          * If doing SELECT INTO, initialize the "into" relation.  We must wait
 763          * till now so we have the "clean" result tuple type to create the new
 764          * table from.
 765          *
 766          * If EXPLAIN, skip creating the "into" relation.
 767          */
 768         intoRelationDesc = NULL;
 769
 770         if (do_select_into && !explainOnly)
 771         {
 772                 char       *intoName;
 773                 Oid                     namespaceId;
 774                 AclResult       aclresult;
 775                 Oid                     intoRelationId;
 776                 TupleDesc       tupdesc;
 777
 778                 /*
 779                  * find namespace to create in, check permissions
 780                  */
 781                 intoName = parseTree->into->relname;
 782                 namespaceId = RangeVarGetCreationNamespace(parseTree->into);
 783
 784                 aclresult = pg_namespace_aclcheck(namespaceId, GetUserId(),
 785                                                                                   ACL_CREATE);
 786                 if (aclresult != ACLCHECK_OK)
 787                         aclcheck_error(aclresult, ACL_KIND_NAMESPACE,
 788                                                    get_namespace_name(namespaceId));
 789
 790                 /*
 791                  * have to copy tupType to get rid of constraints
 792                  */
 793                 tupdesc = CreateTupleDescCopy(tupType);
 794
 795                 intoRelationId = heap_create_with_catalog(intoName,
 796                                                                                                   namespaceId,
 797                                                                                                   tupdesc,
 798                                                                                                   RELKIND_RELATION,
 799                                                                                                   false,
 800                                                                                                   ONCOMMIT_NOOP,
 801                                                                                                   allowSystemTableMods);
 802
 803                 FreeTupleDesc(tupdesc);
 804
 805                 /*
 806                  * Advance command counter so that the newly-created relation's
 807                  * catalog tuples will be visible to heap_open.
 808                  */
 809                 CommandCounterIncrement();
 810
 811                 /*
 812                  * If necessary, create a TOAST table for the into relation. Note
 813                  * that AlterTableCreateToastTable ends with
 814                  * CommandCounterIncrement(), so that the TOAST table will be
 815                  * visible for insertion.
 816                  */
 817                 AlterTableCreateToastTable(intoRelationId, true);
 818
 819                 /*
 820                  * And open the constructed table for writing.
 821                  */
 822                 intoRelationDesc = heap_open(intoRelationId, AccessExclusiveLock);
 823         }
 824
 825         estate->es_into_relation_descriptor = intoRelationDesc;
 826
 827         queryDesc->tupDesc = tupType;
 828         queryDesc->planstate = planstate;
 829 }
 830
 831 /*
 832  * Initialize ResultRelInfo data for one result relation
 833  */
 834 static void
 835 initResultRelInfo(ResultRelInfo *resultRelInfo,
 836                                   Index resultRelationIndex,
 837                                   List *rangeTable,
 838                                   CmdType operation)
 839 {
 840         Oid                     resultRelationOid;
 841         Relation        resultRelationDesc;
 842
 843         resultRelationOid = getrelid(resultRelationIndex, rangeTable);
 844         resultRelationDesc = heap_open(resultRelationOid, RowExclusiveLock);
 845
 846         switch (resultRelationDesc->rd_rel->relkind)
 847         {
 848                 case RELKIND_SEQUENCE:
 849                         ereport(ERROR,
 850                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 851                                          errmsg("cannot change sequence \"%s\"",
 852                                                   RelationGetRelationName(resultRelationDesc))));
 853                         break;
 854                 case RELKIND_TOASTVALUE:
 855                         ereport(ERROR,
 856                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 857                                          errmsg("cannot change TOAST relation \"%s\"",
 858                                                   RelationGetRelationName(resultRelationDesc))));
 859                         break;
 860                 case RELKIND_VIEW:
 861                         ereport(ERROR,
 862                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 863                                          errmsg("cannot change view \"%s\"",
 864                                                   RelationGetRelationName(resultRelationDesc))));
 865                         break;
 866         }
 867
 868         MemSet(resultRelInfo, 0, sizeof(ResultRelInfo));
 869         resultRelInfo->type = T_ResultRelInfo;
 870         resultRelInfo->ri_RangeTableIndex = resultRelationIndex;
 871         resultRelInfo->ri_RelationDesc = resultRelationDesc;
 872         resultRelInfo->ri_NumIndices = 0;
 873         resultRelInfo->ri_IndexRelationDescs = NULL;
 874         resultRelInfo->ri_IndexRelationInfo = NULL;
 875         /* make a copy so as not to depend on relcache info not changing... */
 876         resultRelInfo->ri_TrigDesc = CopyTriggerDesc(resultRelationDesc->trigdesc);
 877         resultRelInfo->ri_TrigFunctions = NULL;
 878         resultRelInfo->ri_ConstraintExprs = NULL;
 879         resultRelInfo->ri_junkFilter = NULL;
 880
 881         /*
 882          * If there are indices on the result relation, open them and save
 883          * descriptors in the result relation info, so that we can add new
 884          * index entries for the tuples we add/update.  We need not do this
 885          * for a DELETE, however, since deletion doesn't affect indexes.
 886          */
 887         if (resultRelationDesc->rd_rel->relhasindex &&
 888                 operation != CMD_DELETE)
 889                 ExecOpenIndices(resultRelInfo);
 890 }
 891
 892 /*
 893  *              ExecContextForcesOids
 894  *
 895  * This is pretty grotty: when doing INSERT, UPDATE, or SELECT INTO,
 896  * we need to ensure that result tuples have space for an OID iff they are
 897  * going to be stored into a relation that has OIDs.  In other contexts
 898  * we are free to choose whether to leave space for OIDs in result tuples
 899  * (we generally don't want to, but we do if a physical-tlist optimization
 900  * is possible).  This routine checks the plan context and returns TRUE if the
 901  * choice is forced, FALSE if the choice is not forced.  In the TRUE case,
 902  * *hasoids is set to the required value.
 903  *
 904  * One reason this is ugly is that all plan nodes in the plan tree will emit
 905  * tuples with space for an OID, though we really only need the topmost node
 906  * to do so.  However, node types like Sort don't project new tuples but just
 907  * return their inputs, and in those cases the requirement propagates down
 908  * to the input node.  Eventually we might make this code smart enough to
 909  * recognize how far down the requirement really goes, but for now we just
 910  * make all plan nodes do the same thing if the top level forces the choice.
 911  *
 912  * We assume that estate->es_result_relation_info is already set up to
 913  * describe the target relation.  Note that in an UPDATE that spans an
 914  * inheritance tree, some of the target relations may have OIDs and some not.
 915  * We have to make the decisions on a per-relation basis as we initialize
 916  * each of the child plans of the topmost Append plan.
 917  *
 918  * SELECT INTO is even uglier, because we don't have the INTO relation's
 919  * descriptor available when this code runs; we have to look aside at a
 920  * flag set by InitPlan().
 921  */
 922 bool
 923 ExecContextForcesOids(PlanState *planstate, bool *hasoids)
 924 {
 925         if (planstate->state->es_select_into)
 926         {
 927                 *hasoids = planstate->state->es_into_oids;
 928                 return true;
 929         }
 930         else
 931         {
 932                 ResultRelInfo *ri = planstate->state->es_result_relation_info;
 933
 934                 if (ri != NULL)
 935                 {
 936                         Relation        rel = ri->ri_RelationDesc;
 937
 938                         if (rel != NULL)
 939                         {
 940                                 *hasoids = rel->rd_rel->relhasoids;
 941                                 return true;
 942                         }
 943                 }
 944         }
 945
 946         return false;
 947 }
 948
 949 /* ----------------------------------------------------------------
 950  *              ExecEndPlan
 951  *
 952  *              Cleans up the query plan -- closes files and frees up storage
 953  *
 954  * NOTE: we are no longer very worried about freeing storage per se
 955  * in this code; FreeExecutorState should be guaranteed to release all
 956  * memory that needs to be released.  What we are worried about doing
 957  * is closing relations and dropping buffer pins.  Thus, for example,
 958  * tuple tables must be cleared or dropped to ensure pins are released.
 959  * ----------------------------------------------------------------
 960  */
 961 void
 962 ExecEndPlan(PlanState *planstate, EState *estate)
 963 {
 964         ResultRelInfo *resultRelInfo;
 965         int                     i;
 966         List       *l;
 967
 968         /*
 969          * shut down any PlanQual processing we were doing
 970          */
 971         if (estate->es_evalPlanQual != NULL)
 972                 EndEvalPlanQual(estate);
 973
 974         /*
 975          * shut down the node-type-specific query processing
 976          */
 977         ExecEndNode(planstate);
 978
 979         /*
 980          * destroy the executor "tuple" table.
 981          */
 982         ExecDropTupleTable(estate->es_tupleTable, true);
 983         estate->es_tupleTable = NULL;
 984
 985         /*
 986          * close the result relation(s) if any, but hold locks until xact
 987          * commit.
 988          */
 989         resultRelInfo = estate->es_result_relations;
 990         for (i = estate->es_num_result_relations; i > 0; i--)
 991         {
 992                 /* Close indices and then the relation itself */
 993                 ExecCloseIndices(resultRelInfo);
 994                 heap_close(resultRelInfo->ri_RelationDesc, NoLock);
 995                 resultRelInfo++;
 996         }
 997
 998         /*
 999          * close the "into" relation if necessary, again keeping lock
1000          */
1001         if (estate->es_into_relation_descriptor != NULL)
1002                 heap_close(estate->es_into_relation_descriptor, NoLock);
1003
1004         /*
1005          * close any relations selected FOR UPDATE, again keeping locks
1006          */
1007         foreach(l, estate->es_rowMark)
1008         {
1009                 execRowMark *erm = lfirst(l);
1010
1011                 heap_close(erm->relation, NoLock);
1012         }
1013 }
1014
1015 /* ----------------------------------------------------------------
1016  *              ExecutePlan
1017  *
1018  *              processes the query plan to retrieve 'numberTuples' tuples in the
1019  *              direction specified.
1020  *
1021  *              Retrieves all tuples if numberTuples is 0
1022  *
1023  *              result is either a slot containing the last tuple in the case
1024  *              of a SELECT or NULL otherwise.
1025  *
1026  * Note: the ctid attribute is a 'junk' attribute that is removed before the
1027  * user can see it
1028  * ----------------------------------------------------------------
1029  */
1030 static TupleTableSlot *
1031 ExecutePlan(EState *estate,
1032                         PlanState *planstate,
1033                         CmdType operation,
1034                         long numberTuples,
1035                         ScanDirection direction,
1036                         DestReceiver *dest)
1037 {
1038         JunkFilter *junkfilter;
1039         TupleTableSlot *slot;
1040         ItemPointer tupleid = NULL;
1041         ItemPointerData tuple_ctid;
1042         long            current_tuple_count;
1043         TupleTableSlot *result;
1044
1045         /*
1046          * initialize local variables
1047          */
1048         slot = NULL;
1049         current_tuple_count = 0;
1050         result = NULL;
1051
1052         /*
1053          * Set the direction.
1054          */
1055         estate->es_direction = direction;
1056
1057         /*
1058          * Process BEFORE EACH STATEMENT triggers
1059          */
1060         switch (operation)
1061         {
1062                 case CMD_UPDATE:
1063                         ExecBSUpdateTriggers(estate, estate->es_result_relation_info);
1064                         break;
1065                 case CMD_DELETE:
1066                         ExecBSDeleteTriggers(estate, estate->es_result_relation_info);
1067                         break;
1068                 case CMD_INSERT:
1069                         ExecBSInsertTriggers(estate, estate->es_result_relation_info);
1070                         break;
1071                 default:
1072                         /* do nothing */
1073                         break;
1074         }
1075
1076         /*
1077          * Loop until we've processed the proper number of tuples from the
1078          * plan.
1079          */
1080
1081         for (;;)
1082         {
1083                 /* Reset the per-output-tuple exprcontext */
1084                 ResetPerTupleExprContext(estate);
1085
1086                 /*
1087                  * Execute the plan and obtain a tuple
1088                  */
1089 lnext:  ;
1090                 if (estate->es_useEvalPlan)
1091                 {
1092                         slot = EvalPlanQualNext(estate);
1093                         if (TupIsNull(slot))
1094                                 slot = ExecProcNode(planstate);
1095                 }
1096                 else
1097                         slot = ExecProcNode(planstate);
1098
1099                 /*
1100                  * if the tuple is null, then we assume there is nothing more to
1101                  * process so we just return null...
1102                  */
1103                 if (TupIsNull(slot))
1104                 {
1105                         result = NULL;
1106                         break;
1107                 }
1108
1109                 /*
1110                  * if we have a junk filter, then project a new tuple with the
1111                  * junk removed.
1112                  *
1113                  * Store this new "clean" tuple in the junkfilter's resultSlot.
1114                  * (Formerly, we stored it back over the "dirty" tuple, which is
1115                  * WRONG because that tuple slot has the wrong descriptor.)
1116                  *
1117                  * Also, extract all the junk information we need.
1118                  */
1119                 if ((junkfilter = estate->es_junkFilter) != NULL)
1120                 {
1121                         Datum           datum;
1122                         HeapTuple       newTuple;
1123                         bool            isNull;
1124
1125                         /*
1126                          * extract the 'ctid' junk attribute.
1127                          */
1128                         if (operation == CMD_UPDATE || operation == CMD_DELETE)
1129                         {
1130                                 if (!ExecGetJunkAttribute(junkfilter,
1131                                                                                   slot,
1132                                                                                   "ctid",
1133                                                                                   &datum,
1134                                                                                   &isNull))
1135                                         elog(ERROR, "could not find junk ctid column");
1136
1137                                 /* shouldn't ever get a null result... */
1138                                 if (isNull)
1139                                         elog(ERROR, "ctid is NULL");
1140
1141                                 tupleid = (ItemPointer) DatumGetPointer(datum);
1142                                 tuple_ctid = *tupleid;  /* make sure we don't free the
1143                                                                                  * ctid!! */
1144                                 tupleid = &tuple_ctid;
1145                         }
1146                         else if (estate->es_rowMark != NIL)
1147                         {
1148                                 List       *l;
1149
1150                 lmark:  ;
1151                                 foreach(l, estate->es_rowMark)
1152                                 {
1153                                         execRowMark *erm = lfirst(l);
1154                                         Buffer          buffer;
1155                                         HeapTupleData tuple;
1156                                         TupleTableSlot *newSlot;
1157                                         int                     test;
1158
1159                                         if (!ExecGetJunkAttribute(junkfilter,
1160                                                                                           slot,
1161                                                                                           erm->resname,
1162                                                                                           &datum,
1163                                                                                           &isNull))
1164                                                 elog(ERROR, "could not find junk \"%s\" column",
1165                                                          erm->resname);
1166
1167                                         /* shouldn't ever get a null result... */
1168                                         if (isNull)
1169                                                 elog(ERROR, "\"%s\" is NULL", erm->resname);
1170
1171                                         tuple.t_self = *((ItemPointer) DatumGetPointer(datum));
1172                                         test = heap_mark4update(erm->relation, &tuple, &buffer,
1173                                                                                         estate->es_snapshot->curcid);
1174                                         ReleaseBuffer(buffer);
1175                                         switch (test)
1176                                         {
1177                                                 case HeapTupleSelfUpdated:
1178                                                         /* treat it as deleted; do not process */
1179                                                         goto lnext;
1180
1181                                                 case HeapTupleMayBeUpdated:
1182                                                         break;
1183
1184                                                 case HeapTupleUpdated:
1185                                                         if (IsXactIsoLevelSerializable)
1186                                                                 ereport(ERROR,
1187                                                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1188                                                                                  errmsg("could not serialize access due to concurrent update")));
1189                                                         if (!(ItemPointerEquals(&(tuple.t_self),
1190                                                                   (ItemPointer) DatumGetPointer(datum))))
1191                                                         {
1192                                                                 newSlot = EvalPlanQual(estate, erm->rti, &(tuple.t_self));
1193                                                                 if (!(TupIsNull(newSlot)))
1194                                                                 {
1195                                                                         slot = newSlot;
1196                                                                         estate->es_useEvalPlan = true;
1197                                                                         goto lmark;
1198                                                                 }
1199                                                         }
1200
1201                                                         /*
1202                                                          * if tuple was deleted or PlanQual failed for
1203                                                          * updated tuple - we must not return this
1204                                                          * tuple!
1205                                                          */
1206                                                         goto lnext;
1207
1208                                                 default:
1209                                                         elog(ERROR, "unrecognized heap_mark4update status: %u",
1210                                                                  test);
1211                                                         return (NULL);
1212                                         }
1213                                 }
1214                         }
1215
1216                         /*
1217                          * Finally create a new "clean" tuple with all junk attributes
1218                          * removed
1219                          */
1220                         newTuple = ExecRemoveJunk(junkfilter, slot);
1221
1222                         slot = ExecStoreTuple(newTuple,         /* tuple to store */
1223                                                                   junkfilter->jf_resultSlot,    /* dest slot */
1224                                                                   InvalidBuffer,                /* this tuple has no
1225                                                                                                                  * buffer */
1226                                                                   true);                /* tuple should be pfreed */
1227                 }
1228
1229                 /*
1230                  * now that we have a tuple, do the appropriate thing with it..
1231                  * either return it to the user, add it to a relation someplace,
1232                  * delete it from a relation, or modify some of its attributes.
1233                  */
1234                 switch (operation)
1235                 {
1236                         case CMD_SELECT:
1237                                 ExecSelect(slot,        /* slot containing tuple */
1238                                                    dest,        /* destination's tuple-receiver obj */
1239                                                    estate);
1240                                 result = slot;
1241                                 break;
1242
1243                         case CMD_INSERT:
1244                                 ExecInsert(slot, tupleid, estate);
1245                                 result = NULL;
1246                                 break;
1247
1248                         case CMD_DELETE:
1249                                 ExecDelete(slot, tupleid, estate);
1250                                 result = NULL;
1251                                 break;
1252
1253                         case CMD_UPDATE:
1254                                 ExecUpdate(slot, tupleid, estate);
1255                                 result = NULL;
1256                                 break;
1257
1258                         default:
1259                                 elog(ERROR, "unrecognized operation code: %d",
1260                                          (int) operation);
1261                                 result = NULL;
1262                                 break;
1263                 }
1264
1265                 /*
1266                  * check our tuple count.. if we've processed the proper number
1267                  * then quit, else loop again and process more tuples.  Zero
1268                  * numberTuples means no limit.
1269                  */
1270                 current_tuple_count++;
1271                 if (numberTuples && numberTuples == current_tuple_count)
1272                         break;
1273         }
1274
1275         /*
1276          * Process AFTER EACH STATEMENT triggers
1277          */
1278         switch (operation)
1279         {
1280                 case CMD_UPDATE:
1281                         ExecASUpdateTriggers(estate, estate->es_result_relation_info);
1282                         break;
1283                 case CMD_DELETE:
1284                         ExecASDeleteTriggers(estate, estate->es_result_relation_info);
1285                         break;
1286                 case CMD_INSERT:
1287                         ExecASInsertTriggers(estate, estate->es_result_relation_info);
1288                         break;
1289                 default:
1290                         /* do nothing */
1291                         break;
1292         }
1293
1294         /*
1295          * here, result is either a slot containing a tuple in the case of a
1296          * SELECT or NULL otherwise.
1297          */
1298         return result;
1299 }
1300
1301 /* ----------------------------------------------------------------
1302  *              ExecSelect
1303  *
1304  *              SELECTs are easy.. we just pass the tuple to the appropriate
1305  *              print function.  The only complexity is when we do a
1306  *              "SELECT INTO", in which case we insert the tuple into
1307  *              the appropriate relation (note: this is a newly created relation
1308  *              so we don't need to worry about indices or locks.)
1309  * ----------------------------------------------------------------
1310  */
1311 static void
1312 ExecSelect(TupleTableSlot *slot,
1313                    DestReceiver *dest,
1314                    EState *estate)
1315 {
1316         HeapTuple       tuple;
1317         TupleDesc       attrtype;
1318
1319         /*
1320          * get the heap tuple out of the tuple table slot
1321          */
1322         tuple = slot->val;
1323         attrtype = slot->ttc_tupleDescriptor;
1324
1325         /*
1326          * insert the tuple into the "into relation"
1327          *
1328          * XXX this probably ought to be replaced by a separate destination
1329          */
1330         if (estate->es_into_relation_descriptor != NULL)
1331         {
1332                 heap_insert(estate->es_into_relation_descriptor, tuple,
1333                                         estate->es_snapshot->curcid);
1334                 IncrAppended();
1335         }
1336
1337         /*
1338          * send the tuple to the destination
1339          */
1340         (*dest->receiveTuple) (tuple, attrtype, dest);
1341         IncrRetrieved();
1342         (estate->es_processed)++;
1343 }
1344
1345 /* ----------------------------------------------------------------
1346  *              ExecInsert
1347  *
1348  *              INSERTs are trickier.. we have to insert the tuple into
1349  *              the base relation and insert appropriate tuples into the
1350  *              index relations.
1351  * ----------------------------------------------------------------
1352  */
1353 static void
1354 ExecInsert(TupleTableSlot *slot,
1355                    ItemPointer tupleid,
1356                    EState *estate)
1357 {
1358         HeapTuple       tuple;
1359         ResultRelInfo *resultRelInfo;
1360         Relation        resultRelationDesc;
1361         int                     numIndices;
1362         Oid                     newId;
1363
1364         /*
1365          * get the heap tuple out of the tuple table slot
1366          */
1367         tuple = slot->val;
1368
1369         /*
1370          * get information on the (current) result relation
1371          */
1372         resultRelInfo = estate->es_result_relation_info;
1373         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1374
1375         /* BEFORE ROW INSERT Triggers */
1376         if (resultRelInfo->ri_TrigDesc &&
1377           resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_INSERT] > 0)
1378         {
1379                 HeapTuple       newtuple;
1380
1381                 newtuple = ExecBRInsertTriggers(estate, resultRelInfo, tuple);
1382
1383                 if (newtuple == NULL)   /* "do nothing" */
1384                         return;
1385
1386                 if (newtuple != tuple)  /* modified by Trigger(s) */
1387                 {
1388                         /*
1389                          * Insert modified tuple into tuple table slot, replacing the
1390                          * original.  We assume that it was allocated in per-tuple
1391                          * memory context, and therefore will go away by itself. The
1392                          * tuple table slot should not try to clear it.
1393                          */
1394                         ExecStoreTuple(newtuple, slot, InvalidBuffer, false);
1395                         tuple = newtuple;
1396                 }
1397         }
1398
1399         /*
1400          * Check the constraints of the tuple
1401          */
1402         if (resultRelationDesc->rd_att->constr)
1403                 ExecConstraints(resultRelInfo, slot, estate);
1404
1405         /*
1406          * insert the tuple
1407          */
1408         newId = heap_insert(resultRelationDesc, tuple,
1409                                                 estate->es_snapshot->curcid);
1410
1411         IncrAppended();
1412         (estate->es_processed)++;
1413         estate->es_lastoid = newId;
1414         setLastTid(&(tuple->t_self));
1415
1416         /*
1417          * process indices
1418          *
1419          * Note: heap_insert adds a new tuple to a relation.  As a side effect,
1420          * the tupleid of the new tuple is placed in the new tuple's t_ctid
1421          * field.
1422          */
1423         numIndices = resultRelInfo->ri_NumIndices;
1424         if (numIndices > 0)
1425                 ExecInsertIndexTuples(slot, &(tuple->t_self), estate, false);
1426
1427         /* AFTER ROW INSERT Triggers */
1428         ExecARInsertTriggers(estate, resultRelInfo, tuple);
1429 }
1430
1431 /* ----------------------------------------------------------------
1432  *              ExecDelete
1433  *
1434  *              DELETE is like UPDATE, we delete the tuple and its
1435  *              index tuples.
1436  * ----------------------------------------------------------------
1437  */
1438 static void
1439 ExecDelete(TupleTableSlot *slot,
1440                    ItemPointer tupleid,
1441                    EState *estate)
1442 {
1443         ResultRelInfo *resultRelInfo;
1444         Relation        resultRelationDesc;
1445         ItemPointerData ctid;
1446         int                     result;
1447
1448         /*
1449          * get information on the (current) result relation
1450          */
1451         resultRelInfo = estate->es_result_relation_info;
1452         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1453
1454         /* BEFORE ROW DELETE Triggers */
1455         if (resultRelInfo->ri_TrigDesc &&
1456           resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_DELETE] > 0)
1457         {
1458                 bool            dodelete;
1459
1460                 dodelete = ExecBRDeleteTriggers(estate, resultRelInfo, tupleid,
1461                                                                                 estate->es_snapshot->curcid);
1462
1463                 if (!dodelete)                  /* "do nothing" */
1464                         return;
1465         }
1466
1467         /*
1468          * delete the tuple
1469          */
1470 ldelete:;
1471         result = heap_delete(resultRelationDesc, tupleid,
1472                                                  &ctid,
1473                                                  estate->es_snapshot->curcid,
1474                                                  estate->es_crosscheck_snapshot,
1475                                                  true /* wait for commit */);
1476         switch (result)
1477         {
1478                 case HeapTupleSelfUpdated:
1479                         /* already deleted by self; nothing to do */
1480                         return;
1481
1482                 case HeapTupleMayBeUpdated:
1483                         break;
1484
1485                 case HeapTupleUpdated:
1486                         if (IsXactIsoLevelSerializable)
1487                                 ereport(ERROR,
1488                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1489                                                  errmsg("could not serialize access due to concurrent update")));
1490                         else if (!(ItemPointerEquals(tupleid, &ctid)))
1491                         {
1492                                 TupleTableSlot *epqslot = EvalPlanQual(estate,
1493                                                            resultRelInfo->ri_RangeTableIndex, &ctid);
1494
1495                                 if (!TupIsNull(epqslot))
1496                                 {
1497                                         *tupleid = ctid;
1498                                         goto ldelete;
1499                                 }
1500                         }
1501                         /* tuple already deleted; nothing to do */
1502                         return;
1503
1504                 default:
1505                         elog(ERROR, "unrecognized heap_delete status: %u", result);
1506                         return;
1507         }
1508
1509         IncrDeleted();
1510         (estate->es_processed)++;
1511
1512         /*
1513          * Note: Normally one would think that we have to delete index tuples
1514          * associated with the heap tuple now..
1515          *
1516          * ... but in POSTGRES, we have no need to do this because the vacuum
1517          * daemon automatically opens an index scan and deletes index tuples
1518          * when it finds deleted heap tuples. -cim 9/27/89
1519          */
1520
1521         /* AFTER ROW DELETE Triggers */
1522         ExecARDeleteTriggers(estate, resultRelInfo, tupleid);
1523 }
1524
1525 /* ----------------------------------------------------------------
1526  *              ExecUpdate
1527  *
1528  *              note: we can't run UPDATE queries with transactions
1529  *              off because UPDATEs are actually INSERTs and our
1530  *              scan will mistakenly loop forever, updating the tuple
1531  *              it just inserted..      This should be fixed but until it
1532  *              is, we don't want to get stuck in an infinite loop
1533  *              which corrupts your database..
1534  * ----------------------------------------------------------------
1535  */
1536 static void
1537 ExecUpdate(TupleTableSlot *slot,
1538                    ItemPointer tupleid,
1539                    EState *estate)
1540 {
1541         HeapTuple       tuple;
1542         ResultRelInfo *resultRelInfo;
1543         Relation        resultRelationDesc;
1544         ItemPointerData ctid;
1545         int                     result;
1546         int                     numIndices;
1547
1548         /*
1549          * abort the operation if not running transactions
1550          */
1551         if (IsBootstrapProcessingMode())
1552                 elog(ERROR, "cannot UPDATE during bootstrap");
1553
1554         /*
1555          * get the heap tuple out of the tuple table slot
1556          */
1557         tuple = slot->val;
1558
1559         /*
1560          * get information on the (current) result relation
1561          */
1562         resultRelInfo = estate->es_result_relation_info;
1563         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1564
1565         /* BEFORE ROW UPDATE Triggers */
1566         if (resultRelInfo->ri_TrigDesc &&
1567           resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_UPDATE] > 0)
1568         {
1569                 HeapTuple       newtuple;
1570
1571                 newtuple = ExecBRUpdateTriggers(estate, resultRelInfo,
1572                                                                                 tupleid, tuple,
1573                                                                                 estate->es_snapshot->curcid);
1574
1575                 if (newtuple == NULL)   /* "do nothing" */
1576                         return;
1577
1578                 if (newtuple != tuple)  /* modified by Trigger(s) */
1579                 {
1580                         /*
1581                          * Insert modified tuple into tuple table slot, replacing the
1582                          * original.  We assume that it was allocated in per-tuple
1583                          * memory context, and therefore will go away by itself. The
1584                          * tuple table slot should not try to clear it.
1585                          */
1586                         ExecStoreTuple(newtuple, slot, InvalidBuffer, false);
1587                         tuple = newtuple;
1588                 }
1589         }
1590
1591         /*
1592          * Check the constraints of the tuple
1593          *
1594          * If we generate a new candidate tuple after EvalPlanQual testing, we
1595          * must loop back here and recheck constraints.  (We don't need to
1596          * redo triggers, however.      If there are any BEFORE triggers then
1597          * trigger.c will have done mark4update to lock the correct tuple, so
1598          * there's no need to do them again.)
1599          */
1600 lreplace:;
1601         if (resultRelationDesc->rd_att->constr)
1602                 ExecConstraints(resultRelInfo, slot, estate);
1603
1604         /*
1605          * replace the heap tuple
1606          */
1607         result = heap_update(resultRelationDesc, tupleid, tuple,
1608                                                  &ctid,
1609                                                  estate->es_snapshot->curcid,
1610                                                  estate->es_crosscheck_snapshot,
1611                                                  true /* wait for commit */);
1612         switch (result)
1613         {
1614                 case HeapTupleSelfUpdated:
1615                         /* already deleted by self; nothing to do */
1616                         return;
1617
1618                 case HeapTupleMayBeUpdated:
1619                         break;
1620
1621                 case HeapTupleUpdated:
1622                         if (IsXactIsoLevelSerializable)
1623                                 ereport(ERROR,
1624                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1625                                                  errmsg("could not serialize access due to concurrent update")));
1626                         else if (!(ItemPointerEquals(tupleid, &ctid)))
1627                         {
1628                                 TupleTableSlot *epqslot = EvalPlanQual(estate,
1629                                                            resultRelInfo->ri_RangeTableIndex, &ctid);
1630
1631                                 if (!TupIsNull(epqslot))
1632                                 {
1633                                         *tupleid = ctid;
1634                                         tuple = ExecRemoveJunk(estate->es_junkFilter, epqslot);
1635                                         slot = ExecStoreTuple(tuple,
1636                                                                         estate->es_junkFilter->jf_resultSlot,
1637                                                                                   InvalidBuffer, true);
1638                                         goto lreplace;
1639                                 }
1640                         }
1641                         /* tuple already deleted; nothing to do */
1642                         return;
1643
1644                 default:
1645                         elog(ERROR, "unrecognized heap_update status: %u", result);
1646                         return;
1647         }
1648
1649         IncrReplaced();
1650         (estate->es_processed)++;
1651
1652         /*
1653          * Note: instead of having to update the old index tuples associated
1654          * with the heap tuple, all we do is form and insert new index tuples.
1655          * This is because UPDATEs are actually DELETEs and INSERTs and index
1656          * tuple deletion is done automagically by the vacuum daemon. All we
1657          * do is insert new index tuples.  -cim 9/27/89
1658          */
1659
1660         /*
1661          * process indices
1662          *
1663          * heap_update updates a tuple in the base relation by invalidating it
1664          * and then inserting a new tuple to the relation.      As a side effect,
1665          * the tupleid of the new tuple is placed in the new tuple's t_ctid
1666          * field.  So we now insert index tuples using the new tupleid stored
1667          * there.
1668          */
1669
1670         numIndices = resultRelInfo->ri_NumIndices;
1671         if (numIndices > 0)
1672                 ExecInsertIndexTuples(slot, &(tuple->t_self), estate, false);
1673
1674         /* AFTER ROW UPDATE Triggers */
1675         ExecARUpdateTriggers(estate, resultRelInfo, tupleid, tuple);
1676 }
1677
1678 static const char *
1679 ExecRelCheck(ResultRelInfo *resultRelInfo,
1680                          TupleTableSlot *slot, EState *estate)
1681 {
1682         Relation        rel = resultRelInfo->ri_RelationDesc;
1683         int                     ncheck = rel->rd_att->constr->num_check;
1684         ConstrCheck *check = rel->rd_att->constr->check;
1685         ExprContext *econtext;
1686         MemoryContext oldContext;
1687         List       *qual;
1688         int                     i;
1689
1690         /*
1691          * If first time through for this result relation, build expression
1692          * nodetrees for rel's constraint expressions.  Keep them in the
1693          * per-query memory context so they'll survive throughout the query.
1694          */
1695         if (resultRelInfo->ri_ConstraintExprs == NULL)
1696         {
1697                 oldContext = MemoryContextSwitchTo(estate->es_query_cxt);
1698                 resultRelInfo->ri_ConstraintExprs =
1699                         (List **) palloc(ncheck * sizeof(List *));
1700                 for (i = 0; i < ncheck; i++)
1701                 {
1702                         /* ExecQual wants implicit-AND form */
1703                         qual = make_ands_implicit(stringToNode(check[i].ccbin));
1704                         resultRelInfo->ri_ConstraintExprs[i] = (List *)
1705                                 ExecPrepareExpr((Expr *) qual, estate);
1706                 }
1707                 MemoryContextSwitchTo(oldContext);
1708         }
1709
1710         /*
1711          * We will use the EState's per-tuple context for evaluating
1712          * constraint expressions (creating it if it's not already there).
1713          */
1714         econtext = GetPerTupleExprContext(estate);
1715
1716         /* Arrange for econtext's scan tuple to be the tuple under test */
1717         econtext->ecxt_scantuple = slot;
1718
1719         /* And evaluate the constraints */
1720         for (i = 0; i < ncheck; i++)
1721         {
1722                 qual = resultRelInfo->ri_ConstraintExprs[i];
1723
1724                 /*
1725                  * NOTE: SQL92 specifies that a NULL result from a constraint
1726                  * expression is not to be treated as a failure.  Therefore, tell
1727                  * ExecQual to return TRUE for NULL.
1728                  */
1729                 if (!ExecQual(qual, econtext, true))
1730                         return check[i].ccname;
1731         }
1732
1733         /* NULL result means no error */
1734         return NULL;
1735 }
1736
1737 void
1738 ExecConstraints(ResultRelInfo *resultRelInfo,
1739                                 TupleTableSlot *slot, EState *estate)
1740 {
1741         Relation        rel = resultRelInfo->ri_RelationDesc;
1742         HeapTuple       tuple = slot->val;
1743         TupleConstr *constr = rel->rd_att->constr;
1744
1745         Assert(constr);
1746
1747         if (constr->has_not_null)
1748         {
1749                 int                     natts = rel->rd_att->natts;
1750                 int                     attrChk;
1751
1752                 for (attrChk = 1; attrChk <= natts; attrChk++)
1753                 {
1754                         if (rel->rd_att->attrs[attrChk - 1]->attnotnull &&
1755                                 heap_attisnull(tuple, attrChk))
1756                                 ereport(ERROR,
1757                                                 (errcode(ERRCODE_NOT_NULL_VIOLATION),
1758                                                  errmsg("null value in column \"%s\" violates not-null constraint",
1759                                         NameStr(rel->rd_att->attrs[attrChk - 1]->attname))));
1760                 }
1761         }
1762
1763         if (constr->num_check > 0)
1764         {
1765                 const char *failed;
1766
1767                 if ((failed = ExecRelCheck(resultRelInfo, slot, estate)) != NULL)
1768                         ereport(ERROR,
1769                                         (errcode(ERRCODE_CHECK_VIOLATION),
1770                                          errmsg("new row for relation \"%s\" violates check constraint \"%s\"",
1771                                                         RelationGetRelationName(rel), failed)));
1772         }
1773 }
1774
1775 /*
1776  * Check a modified tuple to see if we want to process its updated version
1777  * under READ COMMITTED rules.
1778  *
1779  * See backend/executor/README for some info about how this works.
1780  */
1781 TupleTableSlot *
1782 EvalPlanQual(EState *estate, Index rti, ItemPointer tid)
1783 {
1784         evalPlanQual *epq;
1785         EState     *epqstate;
1786         Relation        relation;
1787         HeapTupleData tuple;
1788         HeapTuple       copyTuple = NULL;
1789         bool            endNode;
1790
1791         Assert(rti != 0);
1792
1793         /*
1794          * find relation containing target tuple
1795          */
1796         if (estate->es_result_relation_info != NULL &&
1797                 estate->es_result_relation_info->ri_RangeTableIndex == rti)
1798                 relation = estate->es_result_relation_info->ri_RelationDesc;
1799         else
1800         {
1801                 List       *l;
1802
1803                 relation = NULL;
1804                 foreach(l, estate->es_rowMark)
1805                 {
1806                         if (((execRowMark *) lfirst(l))->rti == rti)
1807                         {
1808                                 relation = ((execRowMark *) lfirst(l))->relation;
1809                                 break;
1810                         }
1811                 }
1812                 if (relation == NULL)
1813                         elog(ERROR, "could not find RowMark for RT index %u", rti);
1814         }
1815
1816         /*
1817          * fetch tid tuple
1818          *
1819          * Loop here to deal with updated or busy tuples
1820          */
1821         tuple.t_self = *tid;
1822         for (;;)
1823         {
1824                 Buffer          buffer;
1825
1826                 if (heap_fetch(relation, SnapshotDirty, &tuple, &buffer, false, NULL))
1827                 {
1828                         TransactionId xwait = SnapshotDirty->xmax;
1829
1830                         /* xmin should not be dirty... */
1831                         if (TransactionIdIsValid(SnapshotDirty->xmin))
1832                                 elog(ERROR, "t_xmin is uncommitted in tuple to be updated");
1833
1834                         /*
1835                          * If tuple is being updated by other transaction then we have
1836                          * to wait for its commit/abort.
1837                          */
1838                         if (TransactionIdIsValid(xwait))
1839                         {
1840                                 ReleaseBuffer(buffer);
1841                                 XactLockTableWait(xwait);
1842                                 continue;
1843                         }
1844
1845                         /*
1846                          * We got tuple - now copy it for use by recheck query.
1847                          */
1848                         copyTuple = heap_copytuple(&tuple);
1849                         ReleaseBuffer(buffer);
1850                         break;
1851                 }
1852
1853                 /*
1854                  * Oops! Invalid tuple. Have to check is it updated or deleted.
1855                  * Note that it's possible to get invalid SnapshotDirty->tid if
1856                  * tuple updated by this transaction. Have we to check this ?
1857                  */
1858                 if (ItemPointerIsValid(&(SnapshotDirty->tid)) &&
1859                         !(ItemPointerEquals(&(tuple.t_self), &(SnapshotDirty->tid))))
1860                 {
1861                         /* updated, so look at the updated copy */
1862                         tuple.t_self = SnapshotDirty->tid;
1863                         continue;
1864                 }
1865
1866                 /*
1867                  * Deleted or updated by this transaction; forget it.
1868                  */
1869                 return NULL;
1870         }
1871
1872         /*
1873          * For UPDATE/DELETE we have to return tid of actual row we're
1874          * executing PQ for.
1875          */
1876         *tid = tuple.t_self;
1877
1878         /*
1879          * Need to run a recheck subquery.      Find or create a PQ stack entry.
1880          */
1881         epq = estate->es_evalPlanQual;
1882         endNode = true;
1883
1884         if (epq != NULL && epq->rti == 0)
1885         {
1886                 /* Top PQ stack entry is idle, so re-use it */
1887                 Assert(!(estate->es_useEvalPlan) && epq->next == NULL);
1888                 epq->rti = rti;
1889                 endNode = false;
1890         }
1891
1892         /*
1893          * If this is request for another RTE - Ra, - then we have to check
1894          * wasn't PlanQual requested for Ra already and if so then Ra' row was
1895          * updated again and we have to re-start old execution for Ra and
1896          * forget all what we done after Ra was suspended. Cool? -:))
1897          */
1898         if (epq != NULL && epq->rti != rti &&
1899                 epq->estate->es_evTuple[rti - 1] != NULL)
1900         {
1901                 do
1902                 {
1903                         evalPlanQual *oldepq;
1904
1905                         /* stop execution */
1906                         EvalPlanQualStop(epq);
1907                         /* pop previous PlanQual from the stack */
1908                         oldepq = epq->next;
1909                         Assert(oldepq && oldepq->rti != 0);
1910                         /* push current PQ to freePQ stack */
1911                         oldepq->free = epq;
1912                         epq = oldepq;
1913                         estate->es_evalPlanQual = epq;
1914                 } while (epq->rti != rti);
1915         }
1916
1917         /*
1918          * If we are requested for another RTE then we have to suspend
1919          * execution of current PlanQual and start execution for new one.
1920          */
1921         if (epq == NULL || epq->rti != rti)
1922         {
1923                 /* try to reuse plan used previously */
1924                 evalPlanQual *newepq = (epq != NULL) ? epq->free : NULL;
1925
1926                 if (newepq == NULL)             /* first call or freePQ stack is empty */
1927                 {
1928                         newepq = (evalPlanQual *) palloc0(sizeof(evalPlanQual));
1929                         newepq->free = NULL;
1930                         newepq->estate = NULL;
1931                         newepq->planstate = NULL;
1932                 }
1933                 else
1934                 {
1935                         /* recycle previously used PlanQual */
1936                         Assert(newepq->estate == NULL);
1937                         epq->free = NULL;
1938                 }
1939                 /* push current PQ to the stack */
1940                 newepq->next = epq;
1941                 epq = newepq;
1942                 estate->es_evalPlanQual = epq;
1943                 epq->rti = rti;
1944                 endNode = false;
1945         }
1946
1947         Assert(epq->rti == rti);
1948
1949         /*
1950          * Ok - we're requested for the same RTE.  Unfortunately we still have
1951          * to end and restart execution of the plan, because ExecReScan
1952          * wouldn't ensure that upper plan nodes would reset themselves.  We
1953          * could make that work if insertion of the target tuple were
1954          * integrated with the Param mechanism somehow, so that the upper plan
1955          * nodes know that their children's outputs have changed.
1956          *
1957          * Note that the stack of free evalPlanQual nodes is quite useless at the
1958          * moment, since it only saves us from pallocing/releasing the
1959          * evalPlanQual nodes themselves.  But it will be useful once we
1960          * implement ReScan instead of end/restart for re-using PlanQual
1961          * nodes.
1962          */
1963         if (endNode)
1964         {
1965                 /* stop execution */
1966                 EvalPlanQualStop(epq);
1967         }
1968
1969         /*
1970          * Initialize new recheck query.
1971          *
1972          * Note: if we were re-using PlanQual plans via ExecReScan, we'd need to
1973          * instead copy down changeable state from the top plan (including
1974          * es_result_relation_info, es_junkFilter) and reset locally
1975          * changeable state in the epq (including es_param_exec_vals,
1976          * es_evTupleNull).
1977          */
1978         EvalPlanQualStart(epq, estate, epq->next);
1979
1980         /*
1981          * free old RTE' tuple, if any, and store target tuple where
1982          * relation's scan node will see it
1983          */
1984         epqstate = epq->estate;
1985         if (epqstate->es_evTuple[rti - 1] != NULL)
1986                 heap_freetuple(epqstate->es_evTuple[rti - 1]);
1987         epqstate->es_evTuple[rti - 1] = copyTuple;
1988
1989         return EvalPlanQualNext(estate);
1990 }
1991
1992 static TupleTableSlot *
1993 EvalPlanQualNext(EState *estate)
1994 {
1995         evalPlanQual *epq = estate->es_evalPlanQual;
1996         MemoryContext oldcontext;
1997         TupleTableSlot *slot;
1998
1999         Assert(epq->rti != 0);
2000
2001 lpqnext:;
2002         oldcontext = MemoryContextSwitchTo(epq->estate->es_query_cxt);
2003         slot = ExecProcNode(epq->planstate);
2004         MemoryContextSwitchTo(oldcontext);
2005
2006         /*
2007          * No more tuples for this PQ. Continue previous one.
2008          */
2009         if (TupIsNull(slot))
2010         {
2011                 evalPlanQual *oldepq;
2012
2013                 /* stop execution */
2014                 EvalPlanQualStop(epq);
2015                 /* pop old PQ from the stack */
2016                 oldepq = epq->next;
2017                 if (oldepq == NULL)
2018                 {
2019                         /* this is the first (oldest) PQ - mark as free */
2020                         epq->rti = 0;
2021                         estate->es_useEvalPlan = false;
2022                         /* and continue Query execution */
2023                         return (NULL);
2024                 }
2025                 Assert(oldepq->rti != 0);
2026                 /* push current PQ to freePQ stack */
2027                 oldepq->free = epq;
2028                 epq = oldepq;
2029                 estate->es_evalPlanQual = epq;
2030                 goto lpqnext;
2031         }
2032
2033         return (slot);
2034 }
2035
2036 static void
2037 EndEvalPlanQual(EState *estate)
2038 {
2039         evalPlanQual *epq = estate->es_evalPlanQual;
2040
2041         if (epq->rti == 0)                      /* plans already shutdowned */
2042         {
2043                 Assert(epq->next == NULL);
2044                 return;
2045         }
2046
2047         for (;;)
2048         {
2049                 evalPlanQual *oldepq;
2050
2051                 /* stop execution */
2052                 EvalPlanQualStop(epq);
2053                 /* pop old PQ from the stack */
2054                 oldepq = epq->next;
2055                 if (oldepq == NULL)
2056                 {
2057                         /* this is the first (oldest) PQ - mark as free */
2058                         epq->rti = 0;
2059                         estate->es_useEvalPlan = false;
2060                         break;
2061                 }
2062                 Assert(oldepq->rti != 0);
2063                 /* push current PQ to freePQ stack */
2064                 oldepq->free = epq;
2065                 epq = oldepq;
2066                 estate->es_evalPlanQual = epq;
2067         }
2068 }
2069
2070 /*
2071  * Start execution of one level of PlanQual.
2072  *
2073  * This is a cut-down version of ExecutorStart(): we copy some state from
2074  * the top-level estate rather than initializing it fresh.
2075  */
2076 static void
2077 EvalPlanQualStart(evalPlanQual *epq, EState *estate, evalPlanQual *priorepq)
2078 {
2079         EState     *epqstate;
2080         int                     rtsize;
2081         MemoryContext oldcontext;
2082
2083         rtsize = length(estate->es_range_table);
2084
2085         epq->estate = epqstate = CreateExecutorState();
2086
2087         oldcontext = MemoryContextSwitchTo(epqstate->es_query_cxt);
2088
2089         /*
2090          * The epqstates share the top query's copy of unchanging state such
2091          * as the snapshot, rangetable, result-rel info, and external Param
2092          * info. They need their own copies of local state, including a tuple
2093          * table, es_param_exec_vals, etc.
2094          */
2095         epqstate->es_direction = ForwardScanDirection;
2096         epqstate->es_snapshot = estate->es_snapshot;
2097         epqstate->es_crosscheck_snapshot = estate->es_crosscheck_snapshot;
2098         epqstate->es_range_table = estate->es_range_table;
2099         epqstate->es_result_relations = estate->es_result_relations;
2100         epqstate->es_num_result_relations = estate->es_num_result_relations;
2101         epqstate->es_result_relation_info = estate->es_result_relation_info;
2102         epqstate->es_junkFilter = estate->es_junkFilter;
2103         epqstate->es_into_relation_descriptor = estate->es_into_relation_descriptor;
2104         epqstate->es_param_list_info = estate->es_param_list_info;
2105         if (estate->es_topPlan->nParamExec > 0)
2106                 epqstate->es_param_exec_vals = (ParamExecData *)
2107                         palloc0(estate->es_topPlan->nParamExec * sizeof(ParamExecData));
2108         epqstate->es_rowMark = estate->es_rowMark;
2109         epqstate->es_instrument = estate->es_instrument;
2110         epqstate->es_select_into = estate->es_select_into;
2111         epqstate->es_into_oids = estate->es_into_oids;
2112         epqstate->es_topPlan = estate->es_topPlan;
2113
2114         /*
2115          * Each epqstate must have its own es_evTupleNull state, but all the
2116          * stack entries share es_evTuple state.  This allows sub-rechecks to
2117          * inherit the value being examined by an outer recheck.
2118          */
2119         epqstate->es_evTupleNull = (bool *) palloc0(rtsize * sizeof(bool));
2120         if (priorepq == NULL)
2121                 /* first PQ stack entry */
2122                 epqstate->es_evTuple = (HeapTuple *)
2123                         palloc0(rtsize * sizeof(HeapTuple));
2124         else
2125                 /* later stack entries share the same storage */
2126                 epqstate->es_evTuple = priorepq->estate->es_evTuple;
2127
2128         epqstate->es_tupleTable =
2129                 ExecCreateTupleTable(estate->es_tupleTable->size);
2130
2131         epq->planstate = ExecInitNode(estate->es_topPlan, epqstate);
2132
2133         MemoryContextSwitchTo(oldcontext);
2134 }
2135
2136 /*
2137  * End execution of one level of PlanQual.
2138  *
2139  * This is a cut-down version of ExecutorEnd(); basically we want to do most
2140  * of the normal cleanup, but *not* close result relations (which we are
2141  * just sharing from the outer query).
2142  */
2143 static void
2144 EvalPlanQualStop(evalPlanQual *epq)
2145 {
2146         EState     *epqstate = epq->estate;
2147         MemoryContext oldcontext;
2148
2149         oldcontext = MemoryContextSwitchTo(epqstate->es_query_cxt);
2150
2151         ExecEndNode(epq->planstate);
2152
2153         ExecDropTupleTable(epqstate->es_tupleTable, true);
2154         epqstate->es_tupleTable = NULL;
2155
2156         if (epqstate->es_evTuple[epq->rti - 1] != NULL)
2157         {
2158                 heap_freetuple(epqstate->es_evTuple[epq->rti - 1]);
2159                 epqstate->es_evTuple[epq->rti - 1] = NULL;
2160         }
2161
2162         MemoryContextSwitchTo(oldcontext);
2163
2164         FreeExecutorState(epqstate);
2165
2166         epq->estate = NULL;
2167         epq->planstate = NULL;
2168 }