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  11  *      The old ExecutorMain() has been replaced by ExecutorStart(),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 1.230 2004/03/23 19:35:16 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                                                                                                   true,
 801                                                                                                   0,
 802                                                                                                   ONCOMMIT_NOOP,
 803                                                                                                   allowSystemTableMods);
 804
 805                 FreeTupleDesc(tupdesc);
 806
 807                 /*
 808                  * Advance command counter so that the newly-created relation's
 809                  * catalog tuples will be visible to heap_open.
 810                  */
 811                 CommandCounterIncrement();
 812
 813                 /*
 814                  * If necessary, create a TOAST table for the into relation. Note
 815                  * that AlterTableCreateToastTable ends with
 816                  * CommandCounterIncrement(), so that the TOAST table will be
 817                  * visible for insertion.
 818                  */
 819                 AlterTableCreateToastTable(intoRelationId, true);
 820
 821                 /*
 822                  * And open the constructed table for writing.
 823                  */
 824                 intoRelationDesc = heap_open(intoRelationId, AccessExclusiveLock);
 825         }
 826
 827         estate->es_into_relation_descriptor = intoRelationDesc;
 828
 829         queryDesc->tupDesc = tupType;
 830         queryDesc->planstate = planstate;
 831 }
 832
 833 /*
 834  * Initialize ResultRelInfo data for one result relation
 835  */
 836 static void
 837 initResultRelInfo(ResultRelInfo *resultRelInfo,
 838                                   Index resultRelationIndex,
 839                                   List *rangeTable,
 840                                   CmdType operation)
 841 {
 842         Oid                     resultRelationOid;
 843         Relation        resultRelationDesc;
 844
 845         resultRelationOid = getrelid(resultRelationIndex, rangeTable);
 846         resultRelationDesc = heap_open(resultRelationOid, RowExclusiveLock);
 847
 848         switch (resultRelationDesc->rd_rel->relkind)
 849         {
 850                 case RELKIND_SEQUENCE:
 851                         ereport(ERROR,
 852                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 853                                          errmsg("cannot change sequence \"%s\"",
 854                                                   RelationGetRelationName(resultRelationDesc))));
 855                         break;
 856                 case RELKIND_TOASTVALUE:
 857                         ereport(ERROR,
 858                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 859                                          errmsg("cannot change TOAST relation \"%s\"",
 860                                                   RelationGetRelationName(resultRelationDesc))));
 861                         break;
 862                 case RELKIND_VIEW:
 863                         ereport(ERROR,
 864                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 865                                          errmsg("cannot change view \"%s\"",
 866                                                   RelationGetRelationName(resultRelationDesc))));
 867                         break;
 868         }
 869
 870         MemSet(resultRelInfo, 0, sizeof(ResultRelInfo));
 871         resultRelInfo->type = T_ResultRelInfo;
 872         resultRelInfo->ri_RangeTableIndex = resultRelationIndex;
 873         resultRelInfo->ri_RelationDesc = resultRelationDesc;
 874         resultRelInfo->ri_NumIndices = 0;
 875         resultRelInfo->ri_IndexRelationDescs = NULL;
 876         resultRelInfo->ri_IndexRelationInfo = NULL;
 877         /* make a copy so as not to depend on relcache info not changing... */
 878         resultRelInfo->ri_TrigDesc = CopyTriggerDesc(resultRelationDesc->trigdesc);
 879         resultRelInfo->ri_TrigFunctions = NULL;
 880         resultRelInfo->ri_ConstraintExprs = NULL;
 881         resultRelInfo->ri_junkFilter = NULL;
 882
 883         /*
 884          * If there are indices on the result relation, open them and save
 885          * descriptors in the result relation info, so that we can add new
 886          * index entries for the tuples we add/update.  We need not do this
 887          * for a DELETE, however, since deletion doesn't affect indexes.
 888          */
 889         if (resultRelationDesc->rd_rel->relhasindex &&
 890                 operation != CMD_DELETE)
 891                 ExecOpenIndices(resultRelInfo);
 892 }
 893
 894 /*
 895  *              ExecContextForcesOids
 896  *
 897  * This is pretty grotty: when doing INSERT, UPDATE, or SELECT INTO,
 898  * we need to ensure that result tuples have space for an OID iff they are
 899  * going to be stored into a relation that has OIDs.  In other contexts
 900  * we are free to choose whether to leave space for OIDs in result tuples
 901  * (we generally don't want to, but we do if a physical-tlist optimization
 902  * is possible).  This routine checks the plan context and returns TRUE if the
 903  * choice is forced, FALSE if the choice is not forced.  In the TRUE case,
 904  * *hasoids is set to the required value.
 905  *
 906  * One reason this is ugly is that all plan nodes in the plan tree will emit
 907  * tuples with space for an OID, though we really only need the topmost node
 908  * to do so.  However, node types like Sort don't project new tuples but just
 909  * return their inputs, and in those cases the requirement propagates down
 910  * to the input node.  Eventually we might make this code smart enough to
 911  * recognize how far down the requirement really goes, but for now we just
 912  * make all plan nodes do the same thing if the top level forces the choice.
 913  *
 914  * We assume that estate->es_result_relation_info is already set up to
 915  * describe the target relation.  Note that in an UPDATE that spans an
 916  * inheritance tree, some of the target relations may have OIDs and some not.
 917  * We have to make the decisions on a per-relation basis as we initialize
 918  * each of the child plans of the topmost Append plan.
 919  *
 920  * SELECT INTO is even uglier, because we don't have the INTO relation's
 921  * descriptor available when this code runs; we have to look aside at a
 922  * flag set by InitPlan().
 923  */
 924 bool
 925 ExecContextForcesOids(PlanState *planstate, bool *hasoids)
 926 {
 927         if (planstate->state->es_select_into)
 928         {
 929                 *hasoids = planstate->state->es_into_oids;
 930                 return true;
 931         }
 932         else
 933         {
 934                 ResultRelInfo *ri = planstate->state->es_result_relation_info;
 935
 936                 if (ri != NULL)
 937                 {
 938                         Relation        rel = ri->ri_RelationDesc;
 939
 940                         if (rel != NULL)
 941                         {
 942                                 *hasoids = rel->rd_rel->relhasoids;
 943                                 return true;
 944                         }
 945                 }
 946         }
 947
 948         return false;
 949 }
 950
 951 /* ----------------------------------------------------------------
 952  *              ExecEndPlan
 953  *
 954  *              Cleans up the query plan -- closes files and frees up storage
 955  *
 956  * NOTE: we are no longer very worried about freeing storage per se
 957  * in this code; FreeExecutorState should be guaranteed to release all
 958  * memory that needs to be released.  What we are worried about doing
 959  * is closing relations and dropping buffer pins.  Thus, for example,
 960  * tuple tables must be cleared or dropped to ensure pins are released.
 961  * ----------------------------------------------------------------
 962  */
 963 void
 964 ExecEndPlan(PlanState *planstate, EState *estate)
 965 {
 966         ResultRelInfo *resultRelInfo;
 967         int                     i;
 968         List       *l;
 969
 970         /*
 971          * shut down any PlanQual processing we were doing
 972          */
 973         if (estate->es_evalPlanQual != NULL)
 974                 EndEvalPlanQual(estate);
 975
 976         /*
 977          * shut down the node-type-specific query processing
 978          */
 979         ExecEndNode(planstate);
 980
 981         /*
 982          * destroy the executor "tuple" table.
 983          */
 984         ExecDropTupleTable(estate->es_tupleTable, true);
 985         estate->es_tupleTable = NULL;
 986
 987         /*
 988          * close the result relation(s) if any, but hold locks until xact
 989          * commit.
 990          */
 991         resultRelInfo = estate->es_result_relations;
 992         for (i = estate->es_num_result_relations; i > 0; i--)
 993         {
 994                 /* Close indices and then the relation itself */
 995                 ExecCloseIndices(resultRelInfo);
 996                 heap_close(resultRelInfo->ri_RelationDesc, NoLock);
 997                 resultRelInfo++;
 998         }
 999
1000         /*
1001          * close the "into" relation if necessary, again keeping lock
1002          */
1003         if (estate->es_into_relation_descriptor != NULL)
1004                 heap_close(estate->es_into_relation_descriptor, NoLock);
1005
1006         /*
1007          * close any relations selected FOR UPDATE, again keeping locks
1008          */
1009         foreach(l, estate->es_rowMark)
1010         {
1011                 execRowMark *erm = lfirst(l);
1012
1013                 heap_close(erm->relation, NoLock);
1014         }
1015 }
1016
1017 /* ----------------------------------------------------------------
1018  *              ExecutePlan
1019  *
1020  *              processes the query plan to retrieve 'numberTuples' tuples in the
1021  *              direction specified.
1022  *
1023  *              Retrieves all tuples if numberTuples is 0
1024  *
1025  *              result is either a slot containing the last tuple in the case
1026  *              of a SELECT or NULL otherwise.
1027  *
1028  * Note: the ctid attribute is a 'junk' attribute that is removed before the
1029  * user can see it
1030  * ----------------------------------------------------------------
1031  */
1032 static TupleTableSlot *
1033 ExecutePlan(EState *estate,
1034                         PlanState *planstate,
1035                         CmdType operation,
1036                         long numberTuples,
1037                         ScanDirection direction,
1038                         DestReceiver *dest)
1039 {
1040         JunkFilter *junkfilter;
1041         TupleTableSlot *slot;
1042         ItemPointer tupleid = NULL;
1043         ItemPointerData tuple_ctid;
1044         long            current_tuple_count;
1045         TupleTableSlot *result;
1046
1047         /*
1048          * initialize local variables
1049          */
1050         slot = NULL;
1051         current_tuple_count = 0;
1052         result = NULL;
1053
1054         /*
1055          * Set the direction.
1056          */
1057         estate->es_direction = direction;
1058
1059         /*
1060          * Process BEFORE EACH STATEMENT triggers
1061          */
1062         switch (operation)
1063         {
1064                 case CMD_UPDATE:
1065                         ExecBSUpdateTriggers(estate, estate->es_result_relation_info);
1066                         break;
1067                 case CMD_DELETE:
1068                         ExecBSDeleteTriggers(estate, estate->es_result_relation_info);
1069                         break;
1070                 case CMD_INSERT:
1071                         ExecBSInsertTriggers(estate, estate->es_result_relation_info);
1072                         break;
1073                 default:
1074                         /* do nothing */
1075                         break;
1076         }
1077
1078         /*
1079          * Loop until we've processed the proper number of tuples from the
1080          * plan.
1081          */
1082
1083         for (;;)
1084         {
1085                 /* Reset the per-output-tuple exprcontext */
1086                 ResetPerTupleExprContext(estate);
1087
1088                 /*
1089                  * Execute the plan and obtain a tuple
1090                  */
1091 lnext:  ;
1092                 if (estate->es_useEvalPlan)
1093                 {
1094                         slot = EvalPlanQualNext(estate);
1095                         if (TupIsNull(slot))
1096                                 slot = ExecProcNode(planstate);
1097                 }
1098                 else
1099                         slot = ExecProcNode(planstate);
1100
1101                 /*
1102                  * if the tuple is null, then we assume there is nothing more to
1103                  * process so we just return null...
1104                  */
1105                 if (TupIsNull(slot))
1106                 {
1107                         result = NULL;
1108                         break;
1109                 }
1110
1111                 /*
1112                  * if we have a junk filter, then project a new tuple with the
1113                  * junk removed.
1114                  *
1115                  * Store this new "clean" tuple in the junkfilter's resultSlot.
1116                  * (Formerly, we stored it back over the "dirty" tuple, which is
1117                  * WRONG because that tuple slot has the wrong descriptor.)
1118                  *
1119                  * Also, extract all the junk information we need.
1120                  */
1121                 if ((junkfilter = estate->es_junkFilter) != NULL)
1122                 {
1123                         Datum           datum;
1124                         HeapTuple       newTuple;
1125                         bool            isNull;
1126
1127                         /*
1128                          * extract the 'ctid' junk attribute.
1129                          */
1130                         if (operation == CMD_UPDATE || operation == CMD_DELETE)
1131                         {
1132                                 if (!ExecGetJunkAttribute(junkfilter,
1133                                                                                   slot,
1134                                                                                   "ctid",
1135                                                                                   &datum,
1136                                                                                   &isNull))
1137                                         elog(ERROR, "could not find junk ctid column");
1138
1139                                 /* shouldn't ever get a null result... */
1140                                 if (isNull)
1141                                         elog(ERROR, "ctid is NULL");
1142
1143                                 tupleid = (ItemPointer) DatumGetPointer(datum);
1144                                 tuple_ctid = *tupleid;  /* make sure we don't free the
1145                                                                                  * ctid!! */
1146                                 tupleid = &tuple_ctid;
1147                         }
1148                         else if (estate->es_rowMark != NIL)
1149                         {
1150                                 List       *l;
1151
1152                 lmark:  ;
1153                                 foreach(l, estate->es_rowMark)
1154                                 {
1155                                         execRowMark *erm = lfirst(l);
1156                                         Buffer          buffer;
1157                                         HeapTupleData tuple;
1158                                         TupleTableSlot *newSlot;
1159                                         int                     test;
1160
1161                                         if (!ExecGetJunkAttribute(junkfilter,
1162                                                                                           slot,
1163                                                                                           erm->resname,
1164                                                                                           &datum,
1165                                                                                           &isNull))
1166                                                 elog(ERROR, "could not find junk \"%s\" column",
1167                                                          erm->resname);
1168
1169                                         /* shouldn't ever get a null result... */
1170                                         if (isNull)
1171                                                 elog(ERROR, "\"%s\" is NULL", erm->resname);
1172
1173                                         tuple.t_self = *((ItemPointer) DatumGetPointer(datum));
1174                                         test = heap_mark4update(erm->relation, &tuple, &buffer,
1175                                                                                         estate->es_snapshot->curcid);
1176                                         ReleaseBuffer(buffer);
1177                                         switch (test)
1178                                         {
1179                                                 case HeapTupleSelfUpdated:
1180                                                         /* treat it as deleted; do not process */
1181                                                         goto lnext;
1182
1183                                                 case HeapTupleMayBeUpdated:
1184                                                         break;
1185
1186                                                 case HeapTupleUpdated:
1187                                                         if (IsXactIsoLevelSerializable)
1188                                                                 ereport(ERROR,
1189                                                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1190                                                                                  errmsg("could not serialize access due to concurrent update")));
1191                                                         if (!(ItemPointerEquals(&(tuple.t_self),
1192                                                                   (ItemPointer) DatumGetPointer(datum))))
1193                                                         {
1194                                                                 newSlot = EvalPlanQual(estate, erm->rti, &(tuple.t_self));
1195                                                                 if (!(TupIsNull(newSlot)))
1196                                                                 {
1197                                                                         slot = newSlot;
1198                                                                         estate->es_useEvalPlan = true;
1199                                                                         goto lmark;
1200                                                                 }
1201                                                         }
1202
1203                                                         /*
1204                                                          * if tuple was deleted or PlanQual failed for
1205                                                          * updated tuple - we must not return this
1206                                                          * tuple!
1207                                                          */
1208                                                         goto lnext;
1209
1210                                                 default:
1211                                                         elog(ERROR, "unrecognized heap_mark4update status: %u",
1212                                                                  test);
1213                                                         return (NULL);
1214                                         }
1215                                 }
1216                         }
1217
1218                         /*
1219                          * Finally create a new "clean" tuple with all junk attributes
1220                          * removed
1221                          */
1222                         newTuple = ExecRemoveJunk(junkfilter, slot);
1223
1224                         slot = ExecStoreTuple(newTuple,         /* tuple to store */
1225                                                                   junkfilter->jf_resultSlot,    /* dest slot */
1226                                                                   InvalidBuffer,                /* this tuple has no
1227                                                                                                                  * buffer */
1228                                                                   true);                /* tuple should be pfreed */
1229                 }
1230
1231                 /*
1232                  * now that we have a tuple, do the appropriate thing with it..
1233                  * either return it to the user, add it to a relation someplace,
1234                  * delete it from a relation, or modify some of its attributes.
1235                  */
1236                 switch (operation)
1237                 {
1238                         case CMD_SELECT:
1239                                 ExecSelect(slot,        /* slot containing tuple */
1240                                                    dest,        /* destination's tuple-receiver obj */
1241                                                    estate);
1242                                 result = slot;
1243                                 break;
1244
1245                         case CMD_INSERT:
1246                                 ExecInsert(slot, tupleid, estate);
1247                                 result = NULL;
1248                                 break;
1249
1250                         case CMD_DELETE:
1251                                 ExecDelete(slot, tupleid, estate);
1252                                 result = NULL;
1253                                 break;
1254
1255                         case CMD_UPDATE:
1256                                 ExecUpdate(slot, tupleid, estate);
1257                                 result = NULL;
1258                                 break;
1259
1260                         default:
1261                                 elog(ERROR, "unrecognized operation code: %d",
1262                                          (int) operation);
1263                                 result = NULL;
1264                                 break;
1265                 }
1266
1267                 /*
1268                  * check our tuple count.. if we've processed the proper number
1269                  * then quit, else loop again and process more tuples.  Zero
1270                  * numberTuples means no limit.
1271                  */
1272                 current_tuple_count++;
1273                 if (numberTuples && numberTuples == current_tuple_count)
1274                         break;
1275         }
1276
1277         /*
1278          * Process AFTER EACH STATEMENT triggers
1279          */
1280         switch (operation)
1281         {
1282                 case CMD_UPDATE:
1283                         ExecASUpdateTriggers(estate, estate->es_result_relation_info);
1284                         break;
1285                 case CMD_DELETE:
1286                         ExecASDeleteTriggers(estate, estate->es_result_relation_info);
1287                         break;
1288                 case CMD_INSERT:
1289                         ExecASInsertTriggers(estate, estate->es_result_relation_info);
1290                         break;
1291                 default:
1292                         /* do nothing */
1293                         break;
1294         }
1295
1296         /*
1297          * here, result is either a slot containing a tuple in the case of a
1298          * SELECT or NULL otherwise.
1299          */
1300         return result;
1301 }
1302
1303 /* ----------------------------------------------------------------
1304  *              ExecSelect
1305  *
1306  *              SELECTs are easy.. we just pass the tuple to the appropriate
1307  *              print function.  The only complexity is when we do a
1308  *              "SELECT INTO", in which case we insert the tuple into
1309  *              the appropriate relation (note: this is a newly created relation
1310  *              so we don't need to worry about indices or locks.)
1311  * ----------------------------------------------------------------
1312  */
1313 static void
1314 ExecSelect(TupleTableSlot *slot,
1315                    DestReceiver *dest,
1316                    EState *estate)
1317 {
1318         HeapTuple       tuple;
1319         TupleDesc       attrtype;
1320
1321         /*
1322          * get the heap tuple out of the tuple table slot
1323          */
1324         tuple = slot->val;
1325         attrtype = slot->ttc_tupleDescriptor;
1326
1327         /*
1328          * insert the tuple into the "into relation"
1329          *
1330          * XXX this probably ought to be replaced by a separate destination
1331          */
1332         if (estate->es_into_relation_descriptor != NULL)
1333         {
1334                 heap_insert(estate->es_into_relation_descriptor, tuple,
1335                                         estate->es_snapshot->curcid);
1336                 IncrAppended();
1337         }
1338
1339         /*
1340          * send the tuple to the destination
1341          */
1342         (*dest->receiveTuple) (tuple, attrtype, dest);
1343         IncrRetrieved();
1344         (estate->es_processed)++;
1345 }
1346
1347 /* ----------------------------------------------------------------
1348  *              ExecInsert
1349  *
1350  *              INSERTs are trickier.. we have to insert the tuple into
1351  *              the base relation and insert appropriate tuples into the
1352  *              index relations.
1353  * ----------------------------------------------------------------
1354  */
1355 static void
1356 ExecInsert(TupleTableSlot *slot,
1357                    ItemPointer tupleid,
1358                    EState *estate)
1359 {
1360         HeapTuple       tuple;
1361         ResultRelInfo *resultRelInfo;
1362         Relation        resultRelationDesc;
1363         int                     numIndices;
1364         Oid                     newId;
1365
1366         /*
1367          * get the heap tuple out of the tuple table slot
1368          */
1369         tuple = slot->val;
1370
1371         /*
1372          * get information on the (current) result relation
1373          */
1374         resultRelInfo = estate->es_result_relation_info;
1375         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1376
1377         /* BEFORE ROW INSERT Triggers */
1378         if (resultRelInfo->ri_TrigDesc &&
1379           resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_INSERT] > 0)
1380         {
1381                 HeapTuple       newtuple;
1382
1383                 newtuple = ExecBRInsertTriggers(estate, resultRelInfo, tuple);
1384
1385                 if (newtuple == NULL)   /* "do nothing" */
1386                         return;
1387
1388                 if (newtuple != tuple)  /* modified by Trigger(s) */
1389                 {
1390                         /*
1391                          * Insert modified tuple into tuple table slot, replacing the
1392                          * original.  We assume that it was allocated in per-tuple
1393                          * memory context, and therefore will go away by itself. The
1394                          * tuple table slot should not try to clear it.
1395                          */
1396                         ExecStoreTuple(newtuple, slot, InvalidBuffer, false);
1397                         tuple = newtuple;
1398                 }
1399         }
1400
1401         /*
1402          * Check the constraints of the tuple
1403          */
1404         if (resultRelationDesc->rd_att->constr)
1405                 ExecConstraints(resultRelInfo, slot, estate);
1406
1407         /*
1408          * insert the tuple
1409          */
1410         newId = heap_insert(resultRelationDesc, tuple,
1411                                                 estate->es_snapshot->curcid);
1412
1413         IncrAppended();
1414         (estate->es_processed)++;
1415         estate->es_lastoid = newId;
1416         setLastTid(&(tuple->t_self));
1417
1418         /*
1419          * process indices
1420          *
1421          * Note: heap_insert adds a new tuple to a relation.  As a side effect,
1422          * the tupleid of the new tuple is placed in the new tuple's t_ctid
1423          * field.
1424          */
1425         numIndices = resultRelInfo->ri_NumIndices;
1426         if (numIndices > 0)
1427                 ExecInsertIndexTuples(slot, &(tuple->t_self), estate, false);
1428
1429         /* AFTER ROW INSERT Triggers */
1430         ExecARInsertTriggers(estate, resultRelInfo, tuple);
1431 }
1432
1433 /* ----------------------------------------------------------------
1434  *              ExecDelete
1435  *
1436  *              DELETE is like UPDATE, we delete the tuple and its
1437  *              index tuples.
1438  * ----------------------------------------------------------------
1439  */
1440 static void
1441 ExecDelete(TupleTableSlot *slot,
1442                    ItemPointer tupleid,
1443                    EState *estate)
1444 {
1445         ResultRelInfo *resultRelInfo;
1446         Relation        resultRelationDesc;
1447         ItemPointerData ctid;
1448         int                     result;
1449
1450         /*
1451          * get information on the (current) result relation
1452          */
1453         resultRelInfo = estate->es_result_relation_info;
1454         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1455
1456         /* BEFORE ROW DELETE Triggers */
1457         if (resultRelInfo->ri_TrigDesc &&
1458           resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_DELETE] > 0)
1459         {
1460                 bool            dodelete;
1461
1462                 dodelete = ExecBRDeleteTriggers(estate, resultRelInfo, tupleid,
1463                                                                                 estate->es_snapshot->curcid);
1464
1465                 if (!dodelete)                  /* "do nothing" */
1466                         return;
1467         }
1468
1469         /*
1470          * delete the tuple
1471          */
1472 ldelete:;
1473         result = heap_delete(resultRelationDesc, tupleid,
1474                                                  &ctid,
1475                                                  estate->es_snapshot->curcid,
1476                                                  estate->es_crosscheck_snapshot,
1477                                                  true /* wait for commit */);
1478         switch (result)
1479         {
1480                 case HeapTupleSelfUpdated:
1481                         /* already deleted by self; nothing to do */
1482                         return;
1483
1484                 case HeapTupleMayBeUpdated:
1485                         break;
1486
1487                 case HeapTupleUpdated:
1488                         if (IsXactIsoLevelSerializable)
1489                                 ereport(ERROR,
1490                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1491                                                  errmsg("could not serialize access due to concurrent update")));
1492                         else if (!(ItemPointerEquals(tupleid, &ctid)))
1493                         {
1494                                 TupleTableSlot *epqslot = EvalPlanQual(estate,
1495                                                            resultRelInfo->ri_RangeTableIndex, &ctid);
1496
1497                                 if (!TupIsNull(epqslot))
1498                                 {
1499                                         *tupleid = ctid;
1500                                         goto ldelete;
1501                                 }
1502                         }
1503                         /* tuple already deleted; nothing to do */
1504                         return;
1505
1506                 default:
1507                         elog(ERROR, "unrecognized heap_delete status: %u", result);
1508                         return;
1509         }
1510
1511         IncrDeleted();
1512         (estate->es_processed)++;
1513
1514         /*
1515          * Note: Normally one would think that we have to delete index tuples
1516          * associated with the heap tuple now..
1517          *
1518          * ... but in POSTGRES, we have no need to do this because the vacuum
1519          * daemon automatically opens an index scan and deletes index tuples
1520          * when it finds deleted heap tuples. -cim 9/27/89
1521          */
1522
1523         /* AFTER ROW DELETE Triggers */
1524         ExecARDeleteTriggers(estate, resultRelInfo, tupleid);
1525 }
1526
1527 /* ----------------------------------------------------------------
1528  *              ExecUpdate
1529  *
1530  *              note: we can't run UPDATE queries with transactions
1531  *              off because UPDATEs are actually INSERTs and our
1532  *              scan will mistakenly loop forever, updating the tuple
1533  *              it just inserted..      This should be fixed but until it
1534  *              is, we don't want to get stuck in an infinite loop
1535  *              which corrupts your database..
1536  * ----------------------------------------------------------------
1537  */
1538 static void
1539 ExecUpdate(TupleTableSlot *slot,
1540                    ItemPointer tupleid,
1541                    EState *estate)
1542 {
1543         HeapTuple       tuple;
1544         ResultRelInfo *resultRelInfo;
1545         Relation        resultRelationDesc;
1546         ItemPointerData ctid;
1547         int                     result;
1548         int                     numIndices;
1549
1550         /*
1551          * abort the operation if not running transactions
1552          */
1553         if (IsBootstrapProcessingMode())
1554                 elog(ERROR, "cannot UPDATE during bootstrap");
1555
1556         /*
1557          * get the heap tuple out of the tuple table slot
1558          */
1559         tuple = slot->val;
1560
1561         /*
1562          * get information on the (current) result relation
1563          */
1564         resultRelInfo = estate->es_result_relation_info;
1565         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1566
1567         /* BEFORE ROW UPDATE Triggers */
1568         if (resultRelInfo->ri_TrigDesc &&
1569           resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_UPDATE] > 0)
1570         {
1571                 HeapTuple       newtuple;
1572
1573                 newtuple = ExecBRUpdateTriggers(estate, resultRelInfo,
1574                                                                                 tupleid, tuple,
1575                                                                                 estate->es_snapshot->curcid);
1576
1577                 if (newtuple == NULL)   /* "do nothing" */
1578                         return;
1579
1580                 if (newtuple != tuple)  /* modified by Trigger(s) */
1581                 {
1582                         /*
1583                          * Insert modified tuple into tuple table slot, replacing the
1584                          * original.  We assume that it was allocated in per-tuple
1585                          * memory context, and therefore will go away by itself. The
1586                          * tuple table slot should not try to clear it.
1587                          */
1588                         ExecStoreTuple(newtuple, slot, InvalidBuffer, false);
1589                         tuple = newtuple;
1590                 }
1591         }
1592
1593         /*
1594          * Check the constraints of the tuple
1595          *
1596          * If we generate a new candidate tuple after EvalPlanQual testing, we
1597          * must loop back here and recheck constraints.  (We don't need to
1598          * redo triggers, however.      If there are any BEFORE triggers then
1599          * trigger.c will have done mark4update to lock the correct tuple, so
1600          * there's no need to do them again.)
1601          */
1602 lreplace:;
1603         if (resultRelationDesc->rd_att->constr)
1604                 ExecConstraints(resultRelInfo, slot, estate);
1605
1606         /*
1607          * replace the heap tuple
1608          */
1609         result = heap_update(resultRelationDesc, tupleid, tuple,
1610                                                  &ctid,
1611                                                  estate->es_snapshot->curcid,
1612                                                  estate->es_crosscheck_snapshot,
1613                                                  true /* wait for commit */);
1614         switch (result)
1615         {
1616                 case HeapTupleSelfUpdated:
1617                         /* already deleted by self; nothing to do */
1618                         return;
1619
1620                 case HeapTupleMayBeUpdated:
1621                         break;
1622
1623                 case HeapTupleUpdated:
1624                         if (IsXactIsoLevelSerializable)
1625                                 ereport(ERROR,
1626                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1627                                                  errmsg("could not serialize access due to concurrent update")));
1628                         else if (!(ItemPointerEquals(tupleid, &ctid)))
1629                         {
1630                                 TupleTableSlot *epqslot = EvalPlanQual(estate,
1631                                                            resultRelInfo->ri_RangeTableIndex, &ctid);
1632
1633                                 if (!TupIsNull(epqslot))
1634                                 {
1635                                         *tupleid = ctid;
1636                                         tuple = ExecRemoveJunk(estate->es_junkFilter, epqslot);
1637                                         slot = ExecStoreTuple(tuple,
1638                                                                         estate->es_junkFilter->jf_resultSlot,
1639                                                                                   InvalidBuffer, true);
1640                                         goto lreplace;
1641                                 }
1642                         }
1643                         /* tuple already deleted; nothing to do */
1644                         return;
1645
1646                 default:
1647                         elog(ERROR, "unrecognized heap_update status: %u", result);
1648                         return;
1649         }
1650
1651         IncrReplaced();
1652         (estate->es_processed)++;
1653
1654         /*
1655          * Note: instead of having to update the old index tuples associated
1656          * with the heap tuple, all we do is form and insert new index tuples.
1657          * This is because UPDATEs are actually DELETEs and INSERTs and index
1658          * tuple deletion is done automagically by the vacuum daemon. All we
1659          * do is insert new index tuples.  -cim 9/27/89
1660          */
1661
1662         /*
1663          * process indices
1664          *
1665          * heap_update updates a tuple in the base relation by invalidating it
1666          * and then inserting a new tuple to the relation.      As a side effect,
1667          * the tupleid of the new tuple is placed in the new tuple's t_ctid
1668          * field.  So we now insert index tuples using the new tupleid stored
1669          * there.
1670          */
1671
1672         numIndices = resultRelInfo->ri_NumIndices;
1673         if (numIndices > 0)
1674                 ExecInsertIndexTuples(slot, &(tuple->t_self), estate, false);
1675
1676         /* AFTER ROW UPDATE Triggers */
1677         ExecARUpdateTriggers(estate, resultRelInfo, tupleid, tuple);
1678 }
1679
1680 static const char *
1681 ExecRelCheck(ResultRelInfo *resultRelInfo,
1682                          TupleTableSlot *slot, EState *estate)
1683 {
1684         Relation        rel = resultRelInfo->ri_RelationDesc;
1685         int                     ncheck = rel->rd_att->constr->num_check;
1686         ConstrCheck *check = rel->rd_att->constr->check;
1687         ExprContext *econtext;
1688         MemoryContext oldContext;
1689         List       *qual;
1690         int                     i;
1691
1692         /*
1693          * If first time through for this result relation, build expression
1694          * nodetrees for rel's constraint expressions.  Keep them in the
1695          * per-query memory context so they'll survive throughout the query.
1696          */
1697         if (resultRelInfo->ri_ConstraintExprs == NULL)
1698         {
1699                 oldContext = MemoryContextSwitchTo(estate->es_query_cxt);
1700                 resultRelInfo->ri_ConstraintExprs =
1701                         (List **) palloc(ncheck * sizeof(List *));
1702                 for (i = 0; i < ncheck; i++)
1703                 {
1704                         /* ExecQual wants implicit-AND form */
1705                         qual = make_ands_implicit(stringToNode(check[i].ccbin));
1706                         resultRelInfo->ri_ConstraintExprs[i] = (List *)
1707                                 ExecPrepareExpr((Expr *) qual, estate);
1708                 }
1709                 MemoryContextSwitchTo(oldContext);
1710         }
1711
1712         /*
1713          * We will use the EState's per-tuple context for evaluating
1714          * constraint expressions (creating it if it's not already there).
1715          */
1716         econtext = GetPerTupleExprContext(estate);
1717
1718         /* Arrange for econtext's scan tuple to be the tuple under test */
1719         econtext->ecxt_scantuple = slot;
1720
1721         /* And evaluate the constraints */
1722         for (i = 0; i < ncheck; i++)
1723         {
1724                 qual = resultRelInfo->ri_ConstraintExprs[i];
1725
1726                 /*
1727                  * NOTE: SQL92 specifies that a NULL result from a constraint
1728                  * expression is not to be treated as a failure.  Therefore, tell
1729                  * ExecQual to return TRUE for NULL.
1730                  */
1731                 if (!ExecQual(qual, econtext, true))
1732                         return check[i].ccname;
1733         }
1734
1735         /* NULL result means no error */
1736         return NULL;
1737 }
1738
1739 void
1740 ExecConstraints(ResultRelInfo *resultRelInfo,
1741                                 TupleTableSlot *slot, EState *estate)
1742 {
1743         Relation        rel = resultRelInfo->ri_RelationDesc;
1744         HeapTuple       tuple = slot->val;
1745         TupleConstr *constr = rel->rd_att->constr;
1746
1747         Assert(constr);
1748
1749         if (constr->has_not_null)
1750         {
1751                 int                     natts = rel->rd_att->natts;
1752                 int                     attrChk;
1753
1754                 for (attrChk = 1; attrChk <= natts; attrChk++)
1755                 {
1756                         if (rel->rd_att->attrs[attrChk - 1]->attnotnull &&
1757                                 heap_attisnull(tuple, attrChk))
1758                                 ereport(ERROR,
1759                                                 (errcode(ERRCODE_NOT_NULL_VIOLATION),
1760                                                  errmsg("null value in column \"%s\" violates not-null constraint",
1761                                         NameStr(rel->rd_att->attrs[attrChk - 1]->attname))));
1762                 }
1763         }
1764
1765         if (constr->num_check > 0)
1766         {
1767                 const char *failed;
1768
1769                 if ((failed = ExecRelCheck(resultRelInfo, slot, estate)) != NULL)
1770                         ereport(ERROR,
1771                                         (errcode(ERRCODE_CHECK_VIOLATION),
1772                                          errmsg("new row for relation \"%s\" violates check constraint \"%s\"",
1773                                                         RelationGetRelationName(rel), failed)));
1774         }
1775 }
1776
1777 /*
1778  * Check a modified tuple to see if we want to process its updated version
1779  * under READ COMMITTED rules.
1780  *
1781  * See backend/executor/README for some info about how this works.
1782  */
1783 TupleTableSlot *
1784 EvalPlanQual(EState *estate, Index rti, ItemPointer tid)
1785 {
1786         evalPlanQual *epq;
1787         EState     *epqstate;
1788         Relation        relation;
1789         HeapTupleData tuple;
1790         HeapTuple       copyTuple = NULL;
1791         bool            endNode;
1792
1793         Assert(rti != 0);
1794
1795         /*
1796          * find relation containing target tuple
1797          */
1798         if (estate->es_result_relation_info != NULL &&
1799                 estate->es_result_relation_info->ri_RangeTableIndex == rti)
1800                 relation = estate->es_result_relation_info->ri_RelationDesc;
1801         else
1802         {
1803                 List       *l;
1804
1805                 relation = NULL;
1806                 foreach(l, estate->es_rowMark)
1807                 {
1808                         if (((execRowMark *) lfirst(l))->rti == rti)
1809                         {
1810                                 relation = ((execRowMark *) lfirst(l))->relation;
1811                                 break;
1812                         }
1813                 }
1814                 if (relation == NULL)
1815                         elog(ERROR, "could not find RowMark for RT index %u", rti);
1816         }
1817
1818         /*
1819          * fetch tid tuple
1820          *
1821          * Loop here to deal with updated or busy tuples
1822          */
1823         tuple.t_self = *tid;
1824         for (;;)
1825         {
1826                 Buffer          buffer;
1827
1828                 if (heap_fetch(relation, SnapshotDirty, &tuple, &buffer, false, NULL))
1829                 {
1830                         TransactionId xwait = SnapshotDirty->xmax;
1831
1832                         /* xmin should not be dirty... */
1833                         if (TransactionIdIsValid(SnapshotDirty->xmin))
1834                                 elog(ERROR, "t_xmin is uncommitted in tuple to be updated");
1835
1836                         /*
1837                          * If tuple is being updated by other transaction then we have
1838                          * to wait for its commit/abort.
1839                          */
1840                         if (TransactionIdIsValid(xwait))
1841                         {
1842                                 ReleaseBuffer(buffer);
1843                                 XactLockTableWait(xwait);
1844                                 continue;
1845                         }
1846
1847                         /*
1848                          * We got tuple - now copy it for use by recheck query.
1849                          */
1850                         copyTuple = heap_copytuple(&tuple);
1851                         ReleaseBuffer(buffer);
1852                         break;
1853                 }
1854
1855                 /*
1856                  * Oops! Invalid tuple. Have to check is it updated or deleted.
1857                  * Note that it's possible to get invalid SnapshotDirty->tid if
1858                  * tuple updated by this transaction. Have we to check this ?
1859                  */
1860                 if (ItemPointerIsValid(&(SnapshotDirty->tid)) &&
1861                         !(ItemPointerEquals(&(tuple.t_self), &(SnapshotDirty->tid))))
1862                 {
1863                         /* updated, so look at the updated copy */
1864                         tuple.t_self = SnapshotDirty->tid;
1865                         continue;
1866                 }
1867
1868                 /*
1869                  * Deleted or updated by this transaction; forget it.
1870                  */
1871                 return NULL;
1872         }
1873
1874         /*
1875          * For UPDATE/DELETE we have to return tid of actual row we're
1876          * executing PQ for.
1877          */
1878         *tid = tuple.t_self;
1879
1880         /*
1881          * Need to run a recheck subquery.      Find or create a PQ stack entry.
1882          */
1883         epq = estate->es_evalPlanQual;
1884         endNode = true;
1885
1886         if (epq != NULL && epq->rti == 0)
1887         {
1888                 /* Top PQ stack entry is idle, so re-use it */
1889                 Assert(!(estate->es_useEvalPlan) && epq->next == NULL);
1890                 epq->rti = rti;
1891                 endNode = false;
1892         }
1893
1894         /*
1895          * If this is request for another RTE - Ra, - then we have to check
1896          * wasn't PlanQual requested for Ra already and if so then Ra' row was
1897          * updated again and we have to re-start old execution for Ra and
1898          * forget all what we done after Ra was suspended. Cool? -:))
1899          */
1900         if (epq != NULL && epq->rti != rti &&
1901                 epq->estate->es_evTuple[rti - 1] != NULL)
1902         {
1903                 do
1904                 {
1905                         evalPlanQual *oldepq;
1906
1907                         /* stop execution */
1908                         EvalPlanQualStop(epq);
1909                         /* pop previous PlanQual from the stack */
1910                         oldepq = epq->next;
1911                         Assert(oldepq && oldepq->rti != 0);
1912                         /* push current PQ to freePQ stack */
1913                         oldepq->free = epq;
1914                         epq = oldepq;
1915                         estate->es_evalPlanQual = epq;
1916                 } while (epq->rti != rti);
1917         }
1918
1919         /*
1920          * If we are requested for another RTE then we have to suspend
1921          * execution of current PlanQual and start execution for new one.
1922          */
1923         if (epq == NULL || epq->rti != rti)
1924         {
1925                 /* try to reuse plan used previously */
1926                 evalPlanQual *newepq = (epq != NULL) ? epq->free : NULL;
1927
1928                 if (newepq == NULL)             /* first call or freePQ stack is empty */
1929                 {
1930                         newepq = (evalPlanQual *) palloc0(sizeof(evalPlanQual));
1931                         newepq->free = NULL;
1932                         newepq->estate = NULL;
1933                         newepq->planstate = NULL;
1934                 }
1935                 else
1936                 {
1937                         /* recycle previously used PlanQual */
1938                         Assert(newepq->estate == NULL);
1939                         epq->free = NULL;
1940                 }
1941                 /* push current PQ to the stack */
1942                 newepq->next = epq;
1943                 epq = newepq;
1944                 estate->es_evalPlanQual = epq;
1945                 epq->rti = rti;
1946                 endNode = false;
1947         }
1948
1949         Assert(epq->rti == rti);
1950
1951         /*
1952          * Ok - we're requested for the same RTE.  Unfortunately we still have
1953          * to end and restart execution of the plan, because ExecReScan
1954          * wouldn't ensure that upper plan nodes would reset themselves.  We
1955          * could make that work if insertion of the target tuple were
1956          * integrated with the Param mechanism somehow, so that the upper plan
1957          * nodes know that their children's outputs have changed.
1958          *
1959          * Note that the stack of free evalPlanQual nodes is quite useless at the
1960          * moment, since it only saves us from pallocing/releasing the
1961          * evalPlanQual nodes themselves.  But it will be useful once we
1962          * implement ReScan instead of end/restart for re-using PlanQual
1963          * nodes.
1964          */
1965         if (endNode)
1966         {
1967                 /* stop execution */
1968                 EvalPlanQualStop(epq);
1969         }
1970
1971         /*
1972          * Initialize new recheck query.
1973          *
1974          * Note: if we were re-using PlanQual plans via ExecReScan, we'd need to
1975          * instead copy down changeable state from the top plan (including
1976          * es_result_relation_info, es_junkFilter) and reset locally
1977          * changeable state in the epq (including es_param_exec_vals,
1978          * es_evTupleNull).
1979          */
1980         EvalPlanQualStart(epq, estate, epq->next);
1981
1982         /*
1983          * free old RTE' tuple, if any, and store target tuple where
1984          * relation's scan node will see it
1985          */
1986         epqstate = epq->estate;
1987         if (epqstate->es_evTuple[rti - 1] != NULL)
1988                 heap_freetuple(epqstate->es_evTuple[rti - 1]);
1989         epqstate->es_evTuple[rti - 1] = copyTuple;
1990
1991         return EvalPlanQualNext(estate);
1992 }
1993
1994 static TupleTableSlot *
1995 EvalPlanQualNext(EState *estate)
1996 {
1997         evalPlanQual *epq = estate->es_evalPlanQual;
1998         MemoryContext oldcontext;
1999         TupleTableSlot *slot;
2000
2001         Assert(epq->rti != 0);
2002
2003 lpqnext:;
2004         oldcontext = MemoryContextSwitchTo(epq->estate->es_query_cxt);
2005         slot = ExecProcNode(epq->planstate);
2006         MemoryContextSwitchTo(oldcontext);
2007
2008         /*
2009          * No more tuples for this PQ. Continue previous one.
2010          */
2011         if (TupIsNull(slot))
2012         {
2013                 evalPlanQual *oldepq;
2014
2015                 /* stop execution */
2016                 EvalPlanQualStop(epq);
2017                 /* pop old PQ from the stack */
2018                 oldepq = epq->next;
2019                 if (oldepq == NULL)
2020                 {
2021                         /* this is the first (oldest) PQ - mark as free */
2022                         epq->rti = 0;
2023                         estate->es_useEvalPlan = false;
2024                         /* and continue Query execution */
2025                         return (NULL);
2026                 }
2027                 Assert(oldepq->rti != 0);
2028                 /* push current PQ to freePQ stack */
2029                 oldepq->free = epq;
2030                 epq = oldepq;
2031                 estate->es_evalPlanQual = epq;
2032                 goto lpqnext;
2033         }
2034
2035         return (slot);
2036 }
2037
2038 static void
2039 EndEvalPlanQual(EState *estate)
2040 {
2041         evalPlanQual *epq = estate->es_evalPlanQual;
2042
2043         if (epq->rti == 0)                      /* plans already shutdowned */
2044         {
2045                 Assert(epq->next == NULL);
2046                 return;
2047         }
2048
2049         for (;;)
2050         {
2051                 evalPlanQual *oldepq;
2052
2053                 /* stop execution */
2054                 EvalPlanQualStop(epq);
2055                 /* pop old PQ from the stack */
2056                 oldepq = epq->next;
2057                 if (oldepq == NULL)
2058                 {
2059                         /* this is the first (oldest) PQ - mark as free */
2060                         epq->rti = 0;
2061                         estate->es_useEvalPlan = false;
2062                         break;
2063                 }
2064                 Assert(oldepq->rti != 0);
2065                 /* push current PQ to freePQ stack */
2066                 oldepq->free = epq;
2067                 epq = oldepq;
2068                 estate->es_evalPlanQual = epq;
2069         }
2070 }
2071
2072 /*
2073  * Start execution of one level of PlanQual.
2074  *
2075  * This is a cut-down version of ExecutorStart(): we copy some state from
2076  * the top-level estate rather than initializing it fresh.
2077  */
2078 static void
2079 EvalPlanQualStart(evalPlanQual *epq, EState *estate, evalPlanQual *priorepq)
2080 {
2081         EState     *epqstate;
2082         int                     rtsize;
2083         MemoryContext oldcontext;
2084
2085         rtsize = length(estate->es_range_table);
2086
2087         epq->estate = epqstate = CreateExecutorState();
2088
2089         oldcontext = MemoryContextSwitchTo(epqstate->es_query_cxt);
2090
2091         /*
2092          * The epqstates share the top query's copy of unchanging state such
2093          * as the snapshot, rangetable, result-rel info, and external Param
2094          * info. They need their own copies of local state, including a tuple
2095          * table, es_param_exec_vals, etc.
2096          */
2097         epqstate->es_direction = ForwardScanDirection;
2098         epqstate->es_snapshot = estate->es_snapshot;
2099         epqstate->es_crosscheck_snapshot = estate->es_crosscheck_snapshot;
2100         epqstate->es_range_table = estate->es_range_table;
2101         epqstate->es_result_relations = estate->es_result_relations;
2102         epqstate->es_num_result_relations = estate->es_num_result_relations;
2103         epqstate->es_result_relation_info = estate->es_result_relation_info;
2104         epqstate->es_junkFilter = estate->es_junkFilter;
2105         epqstate->es_into_relation_descriptor = estate->es_into_relation_descriptor;
2106         epqstate->es_param_list_info = estate->es_param_list_info;
2107         if (estate->es_topPlan->nParamExec > 0)
2108                 epqstate->es_param_exec_vals = (ParamExecData *)
2109                         palloc0(estate->es_topPlan->nParamExec * sizeof(ParamExecData));
2110         epqstate->es_rowMark = estate->es_rowMark;
2111         epqstate->es_instrument = estate->es_instrument;
2112         epqstate->es_select_into = estate->es_select_into;
2113         epqstate->es_into_oids = estate->es_into_oids;
2114         epqstate->es_topPlan = estate->es_topPlan;
2115
2116         /*
2117          * Each epqstate must have its own es_evTupleNull state, but all the
2118          * stack entries share es_evTuple state.  This allows sub-rechecks to
2119          * inherit the value being examined by an outer recheck.
2120          */
2121         epqstate->es_evTupleNull = (bool *) palloc0(rtsize * sizeof(bool));
2122         if (priorepq == NULL)
2123                 /* first PQ stack entry */
2124                 epqstate->es_evTuple = (HeapTuple *)
2125                         palloc0(rtsize * sizeof(HeapTuple));
2126         else
2127                 /* later stack entries share the same storage */
2128                 epqstate->es_evTuple = priorepq->estate->es_evTuple;
2129
2130         epqstate->es_tupleTable =
2131                 ExecCreateTupleTable(estate->es_tupleTable->size);
2132
2133         epq->planstate = ExecInitNode(estate->es_topPlan, epqstate);
2134
2135         MemoryContextSwitchTo(oldcontext);
2136 }
2137
2138 /*
2139  * End execution of one level of PlanQual.
2140  *
2141  * This is a cut-down version of ExecutorEnd(); basically we want to do most
2142  * of the normal cleanup, but *not* close result relations (which we are
2143  * just sharing from the outer query).
2144  */
2145 static void
2146 EvalPlanQualStop(evalPlanQual *epq)
2147 {
2148         EState     *epqstate = epq->estate;
2149         MemoryContext oldcontext;
2150
2151         oldcontext = MemoryContextSwitchTo(epqstate->es_query_cxt);
2152
2153         ExecEndNode(epq->planstate);
2154
2155         ExecDropTupleTable(epqstate->es_tupleTable, true);
2156         epqstate->es_tupleTable = NULL;
2157
2158         if (epqstate->es_evTuple[epq->rti - 1] != NULL)
2159         {
2160                 heap_freetuple(epqstate->es_evTuple[epq->rti - 1]);
2161                 epqstate->es_evTuple[epq->rti - 1] = NULL;
2162         }
2163
2164         MemoryContextSwitchTo(oldcontext);
2165
2166         FreeExecutorState(epqstate);
2167
2168         epq->estate = NULL;
2169         epq->planstate = NULL;
2170 }