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  11  *      The old ExecutorMain() has been replaced by ExecutorStart(),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) 1996-2005, 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 1.255 2005/08/26 03:07:25 tgl Exp $
  30  *
  31  *-------------------------------------------------------------------------
  32  */
  33 #include "postgres.h"
  34
  35 #include "access/heapam.h"
  36 #include "access/xlog.h"
  37 #include "catalog/heap.h"
  38 #include "catalog/namespace.h"
  39 #include "commands/tablecmds.h"
  40 #include "commands/trigger.h"
  41 #include "executor/execdebug.h"
  42 #include "executor/execdefs.h"
  43 #include "executor/instrument.h"
  44 #include "miscadmin.h"
  45 #include "optimizer/clauses.h"
  46 #include "optimizer/var.h"
  47 #include "parser/parsetree.h"
  48 #include "storage/smgr.h"
  49 #include "utils/acl.h"
  50 #include "utils/guc.h"
  51 #include "utils/lsyscache.h"
  52 #include "utils/memutils.h"
  53
  54
  55 typedef struct execRowMark
  56 {
  57         Relation        relation;
  58         Index           rti;
  59         char            resname[32];
  60 } execRowMark;
  61
  62 typedef struct evalPlanQual
  63 {
  64         Index           rti;
  65         EState     *estate;
  66         PlanState  *planstate;
  67         struct evalPlanQual *next;      /* stack of active PlanQual plans */
  68         struct evalPlanQual *free;      /* list of free PlanQual plans */
  69 } evalPlanQual;
  70
  71 /* decls for local routines only used within this module */
  72 static void InitPlan(QueryDesc *queryDesc, bool explainOnly);
  73 static void initResultRelInfo(ResultRelInfo *resultRelInfo,
  74                                   Index resultRelationIndex,
  75                                   List *rangeTable,
  76                                   CmdType operation,
  77                                   bool doInstrument);
  78 static TupleTableSlot *ExecutePlan(EState *estate, PlanState *planstate,
  79                         CmdType operation,
  80                         long numberTuples,
  81                         ScanDirection direction,
  82                         DestReceiver *dest);
  83 static void ExecSelect(TupleTableSlot *slot,
  84                    DestReceiver *dest,
  85                    EState *estate);
  86 static void ExecInsert(TupleTableSlot *slot, ItemPointer tupleid,
  87                    EState *estate);
  88 static void ExecDelete(TupleTableSlot *slot, ItemPointer tupleid,
  89                    EState *estate);
  90 static void ExecUpdate(TupleTableSlot *slot, ItemPointer tupleid,
  91                    EState *estate);
  92 static TupleTableSlot *EvalPlanQualNext(EState *estate);
  93 static void EndEvalPlanQual(EState *estate);
  94 static void ExecCheckRTEPerms(RangeTblEntry *rte);
  95 static void ExecCheckXactReadOnly(Query *parsetree);
  96 static void EvalPlanQualStart(evalPlanQual *epq, EState *estate,
  97                                   evalPlanQual *priorepq);
  98 static void EvalPlanQualStop(evalPlanQual *epq);
  99
 100 /* end of local decls */
 101
 102
 103 /* ----------------------------------------------------------------
 104  *              ExecutorStart
 105  *
 106  *              This routine must be called at the beginning of any execution of any
 107  *              query plan
 108  *
 109  * Takes a QueryDesc previously created by CreateQueryDesc (it's not real
 110  * clear why we bother to separate the two functions, but...).  The tupDesc
 111  * field of the QueryDesc is filled in to describe the tuples that will be
 112  * returned, and the internal fields (estate and planstate) are set up.
 113  *
 114  * If explainOnly is true, we are not actually intending to run the plan,
 115  * only to set up for EXPLAIN; so skip unwanted side-effects.
 116  *
 117  * NB: the CurrentMemoryContext when this is called will become the parent
 118  * of the per-query context used for this Executor invocation.
 119  * ----------------------------------------------------------------
 120  */
 121 void
 122 ExecutorStart(QueryDesc *queryDesc, bool explainOnly)
 123 {
 124         EState     *estate;
 125         MemoryContext oldcontext;
 126
 127         /* sanity checks: queryDesc must not be started already */
 128         Assert(queryDesc != NULL);
 129         Assert(queryDesc->estate == NULL);
 130
 131         /*
 132          * If the transaction is read-only, we need to check if any writes are
 133          * planned to non-temporary tables.
 134          */
 135         if (XactReadOnly && !explainOnly)
 136                 ExecCheckXactReadOnly(queryDesc->parsetree);
 137
 138         /*
 139          * Build EState, switch into per-query memory context for startup.
 140          */
 141         estate = CreateExecutorState();
 142         queryDesc->estate = estate;
 143
 144         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 145
 146         /*
 147          * Fill in parameters, if any, from queryDesc
 148          */
 149         estate->es_param_list_info = queryDesc->params;
 150
 151         if (queryDesc->plantree->nParamExec > 0)
 152                 estate->es_param_exec_vals = (ParamExecData *)
 153                         palloc0(queryDesc->plantree->nParamExec * sizeof(ParamExecData));
 154
 155         /*
 156          * Copy other important information into the EState
 157          */
 158         estate->es_snapshot = queryDesc->snapshot;
 159         estate->es_crosscheck_snapshot = queryDesc->crosscheck_snapshot;
 160         estate->es_instrument = queryDesc->doInstrument;
 161
 162         /*
 163          * Initialize the plan state tree
 164          */
 165         InitPlan(queryDesc, explainOnly);
 166
 167         MemoryContextSwitchTo(oldcontext);
 168 }
 169
 170 /* ----------------------------------------------------------------
 171  *              ExecutorRun
 172  *
 173  *              This is the main routine of the executor module. It accepts
 174  *              the query descriptor from the traffic cop and executes the
 175  *              query plan.
 176  *
 177  *              ExecutorStart must have been called already.
 178  *
 179  *              If direction is NoMovementScanDirection then nothing is done
 180  *              except to start up/shut down the destination.  Otherwise,
 181  *              we retrieve up to 'count' tuples in the specified direction.
 182  *
 183  *              Note: count = 0 is interpreted as no portal limit, i.e., run to
 184  *              completion.
 185  *
 186  * ----------------------------------------------------------------
 187  */
 188 TupleTableSlot *
 189 ExecutorRun(QueryDesc *queryDesc,
 190                         ScanDirection direction, long count)
 191 {
 192         EState     *estate;
 193         CmdType         operation;
 194         DestReceiver *dest;
 195         TupleTableSlot *result;
 196         MemoryContext oldcontext;
 197
 198         /* sanity checks */
 199         Assert(queryDesc != NULL);
 200
 201         estate = queryDesc->estate;
 202
 203         Assert(estate != NULL);
 204
 205         /*
 206          * Switch into per-query memory context
 207          */
 208         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 209
 210         /*
 211          * extract information from the query descriptor and the query
 212          * feature.
 213          */
 214         operation = queryDesc->operation;
 215         dest = queryDesc->dest;
 216
 217         /*
 218          * startup tuple receiver
 219          */
 220         estate->es_processed = 0;
 221         estate->es_lastoid = InvalidOid;
 222
 223         (*dest->rStartup) (dest, operation, queryDesc->tupDesc);
 224
 225         /*
 226          * run plan
 227          */
 228         if (direction == NoMovementScanDirection)
 229                 result = NULL;
 230         else
 231                 result = ExecutePlan(estate,
 232                                                          queryDesc->planstate,
 233                                                          operation,
 234                                                          count,
 235                                                          direction,
 236                                                          dest);
 237
 238         /*
 239          * shutdown receiver
 240          */
 241         (*dest->rShutdown) (dest);
 242
 243         MemoryContextSwitchTo(oldcontext);
 244
 245         return result;
 246 }
 247
 248 /* ----------------------------------------------------------------
 249  *              ExecutorEnd
 250  *
 251  *              This routine must be called at the end of execution of any
 252  *              query plan
 253  * ----------------------------------------------------------------
 254  */
 255 void
 256 ExecutorEnd(QueryDesc *queryDesc)
 257 {
 258         EState     *estate;
 259         MemoryContext oldcontext;
 260
 261         /* sanity checks */
 262         Assert(queryDesc != NULL);
 263
 264         estate = queryDesc->estate;
 265
 266         Assert(estate != NULL);
 267
 268         /*
 269          * Switch into per-query memory context to run ExecEndPlan
 270          */
 271         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 272
 273         ExecEndPlan(queryDesc->planstate, estate);
 274
 275         /*
 276          * Must switch out of context before destroying it
 277          */
 278         MemoryContextSwitchTo(oldcontext);
 279
 280         /*
 281          * Release EState and per-query memory context.  This should release
 282          * everything the executor has allocated.
 283          */
 284         FreeExecutorState(estate);
 285
 286         /* Reset queryDesc fields that no longer point to anything */
 287         queryDesc->tupDesc = NULL;
 288         queryDesc->estate = NULL;
 289         queryDesc->planstate = NULL;
 290 }
 291
 292 /* ----------------------------------------------------------------
 293  *              ExecutorRewind
 294  *
 295  *              This routine may be called on an open queryDesc to rewind it
 296  *              to the start.
 297  * ----------------------------------------------------------------
 298  */
 299 void
 300 ExecutorRewind(QueryDesc *queryDesc)
 301 {
 302         EState     *estate;
 303         MemoryContext oldcontext;
 304
 305         /* sanity checks */
 306         Assert(queryDesc != NULL);
 307
 308         estate = queryDesc->estate;
 309
 310         Assert(estate != NULL);
 311
 312         /* It's probably not sensible to rescan updating queries */
 313         Assert(queryDesc->operation == CMD_SELECT);
 314
 315         /*
 316          * Switch into per-query memory context
 317          */
 318         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 319
 320         /*
 321          * rescan plan
 322          */
 323         ExecReScan(queryDesc->planstate, NULL);
 324
 325         MemoryContextSwitchTo(oldcontext);
 326 }
 327
 328
 329 /*
 330  * ExecCheckRTPerms
 331  *              Check access permissions for all relations listed in a range table.
 332  */
 333 void
 334 ExecCheckRTPerms(List *rangeTable)
 335 {
 336         ListCell   *l;
 337
 338         foreach(l, rangeTable)
 339         {
 340                 RangeTblEntry *rte = lfirst(l);
 341
 342                 ExecCheckRTEPerms(rte);
 343         }
 344 }
 345
 346 /*
 347  * ExecCheckRTEPerms
 348  *              Check access permissions for a single RTE.
 349  */
 350 static void
 351 ExecCheckRTEPerms(RangeTblEntry *rte)
 352 {
 353         AclMode         requiredPerms;
 354         Oid                     relOid;
 355         Oid             userid;
 356
 357         /*
 358          * Only plain-relation RTEs need to be checked here.  Subquery RTEs
 359          * are checked by ExecInitSubqueryScan if the subquery is still a
 360          * separate subquery --- if it's been pulled up into our query level
 361          * then the RTEs are in our rangetable and will be checked here.
 362          * Function RTEs are checked by init_fcache when the function is
 363          * prepared for execution. Join and special RTEs need no checks.
 364          */
 365         if (rte->rtekind != RTE_RELATION)
 366                 return;
 367
 368         /*
 369          * No work if requiredPerms is empty.
 370          */
 371         requiredPerms = rte->requiredPerms;
 372         if (requiredPerms == 0)
 373                 return;
 374
 375         relOid = rte->relid;
 376
 377         /*
 378          * userid to check as: current user unless we have a setuid
 379          * indication.
 380          *
 381          * Note: GetUserId() is presently fast enough that there's no harm in
 382          * calling it separately for each RTE.  If that stops being true, we
 383          * could call it once in ExecCheckRTPerms and pass the userid down
 384          * from there.  But for now, no need for the extra clutter.
 385          */
 386         userid = rte->checkAsUser ? rte->checkAsUser : GetUserId();
 387
 388         /*
 389          * We must have *all* the requiredPerms bits, so use aclmask not
 390          * aclcheck.
 391          */
 392         if (pg_class_aclmask(relOid, userid, requiredPerms, ACLMASK_ALL)
 393                 != requiredPerms)
 394                 aclcheck_error(ACLCHECK_NO_PRIV, ACL_KIND_CLASS,
 395                                            get_rel_name(relOid));
 396 }
 397
 398 /*
 399  * Check that the query does not imply any writes to non-temp tables.
 400  */
 401 static void
 402 ExecCheckXactReadOnly(Query *parsetree)
 403 {
 404         ListCell   *l;
 405
 406         /*
 407          * CREATE TABLE AS or SELECT INTO?
 408          *
 409          * XXX should we allow this if the destination is temp?
 410          */
 411         if (parsetree->into != NULL)
 412                 goto fail;
 413
 414         /* Fail if write permissions are requested on any non-temp table */
 415         foreach(l, parsetree->rtable)
 416         {
 417                 RangeTblEntry *rte = lfirst(l);
 418
 419                 if (rte->rtekind == RTE_SUBQUERY)
 420                 {
 421                         ExecCheckXactReadOnly(rte->subquery);
 422                         continue;
 423                 }
 424
 425                 if (rte->rtekind != RTE_RELATION)
 426                         continue;
 427
 428                 if ((rte->requiredPerms & (~ACL_SELECT)) == 0)
 429                         continue;
 430
 431                 if (isTempNamespace(get_rel_namespace(rte->relid)))
 432                         continue;
 433
 434                 goto fail;
 435         }
 436
 437         return;
 438
 439 fail:
 440         ereport(ERROR,
 441                         (errcode(ERRCODE_READ_ONLY_SQL_TRANSACTION),
 442                          errmsg("transaction is read-only")));
 443 }
 444
 445
 446 /* ----------------------------------------------------------------
 447  *              InitPlan
 448  *
 449  *              Initializes the query plan: open files, allocate storage
 450  *              and start up the rule manager
 451  * ----------------------------------------------------------------
 452  */
 453 static void
 454 InitPlan(QueryDesc *queryDesc, bool explainOnly)
 455 {
 456         CmdType         operation = queryDesc->operation;
 457         Query      *parseTree = queryDesc->parsetree;
 458         Plan       *plan = queryDesc->plantree;
 459         EState     *estate = queryDesc->estate;
 460         PlanState  *planstate;
 461         List       *rangeTable;
 462         Relation        intoRelationDesc;
 463         bool            do_select_into;
 464         TupleDesc       tupType;
 465
 466         /*
 467          * Do permissions checks.  It's sufficient to examine the query's top
 468          * rangetable here --- subplan RTEs will be checked during
 469          * ExecInitSubPlan().
 470          */
 471         ExecCheckRTPerms(parseTree->rtable);
 472
 473         /*
 474          * get information from query descriptor
 475          */
 476         rangeTable = parseTree->rtable;
 477
 478         /*
 479          * initialize the node's execution state
 480          */
 481         estate->es_range_table = rangeTable;
 482
 483         /*
 484          * if there is a result relation, initialize result relation stuff
 485          */
 486         if (parseTree->resultRelation != 0 && operation != CMD_SELECT)
 487         {
 488                 List       *resultRelations = parseTree->resultRelations;
 489                 int                     numResultRelations;
 490                 ResultRelInfo *resultRelInfos;
 491
 492                 if (resultRelations != NIL)
 493                 {
 494                         /*
 495                          * Multiple result relations (due to inheritance)
 496                          * parseTree->resultRelations identifies them all
 497                          */
 498                         ResultRelInfo *resultRelInfo;
 499                         ListCell   *l;
 500
 501                         numResultRelations = list_length(resultRelations);
 502                         resultRelInfos = (ResultRelInfo *)
 503                                 palloc(numResultRelations * sizeof(ResultRelInfo));
 504                         resultRelInfo = resultRelInfos;
 505                         foreach(l, resultRelations)
 506                         {
 507                                 initResultRelInfo(resultRelInfo,
 508                                                                   lfirst_int(l),
 509                                                                   rangeTable,
 510                                                                   operation,
 511                                                                   estate->es_instrument);
 512                                 resultRelInfo++;
 513                         }
 514                 }
 515                 else
 516                 {
 517                         /*
 518                          * Single result relation identified by
 519                          * parseTree->resultRelation
 520                          */
 521                         numResultRelations = 1;
 522                         resultRelInfos = (ResultRelInfo *) palloc(sizeof(ResultRelInfo));
 523                         initResultRelInfo(resultRelInfos,
 524                                                           parseTree->resultRelation,
 525                                                           rangeTable,
 526                                                           operation,
 527                                                           estate->es_instrument);
 528                 }
 529
 530                 estate->es_result_relations = resultRelInfos;
 531                 estate->es_num_result_relations = numResultRelations;
 532                 /* Initialize to first or only result rel */
 533                 estate->es_result_relation_info = resultRelInfos;
 534         }
 535         else
 536         {
 537                 /*
 538                  * if no result relation, then set state appropriately
 539                  */
 540                 estate->es_result_relations = NULL;
 541                 estate->es_num_result_relations = 0;
 542                 estate->es_result_relation_info = NULL;
 543         }
 544
 545         /*
 546          * Detect whether we're doing SELECT INTO.  If so, set the es_into_oids
 547          * flag appropriately so that the plan tree will be initialized with
 548          * the correct tuple descriptors.
 549          */
 550         do_select_into = false;
 551
 552         if (operation == CMD_SELECT && parseTree->into != NULL)
 553         {
 554                 do_select_into = true;
 555                 estate->es_select_into = true;
 556                 estate->es_into_oids = parseTree->intoHasOids;
 557         }
 558
 559         /*
 560          * Have to lock relations selected FOR UPDATE/FOR SHARE
 561          */
 562         estate->es_rowMarks = NIL;
 563         estate->es_forUpdate = parseTree->forUpdate;
 564         estate->es_rowNoWait = parseTree->rowNoWait;
 565         if (parseTree->rowMarks != NIL)
 566         {
 567                 ListCell   *l;
 568
 569                 foreach(l, parseTree->rowMarks)
 570                 {
 571                         Index           rti = lfirst_int(l);
 572                         Oid                     relid = getrelid(rti, rangeTable);
 573                         Relation        relation;
 574                         execRowMark *erm;
 575
 576                         relation = heap_open(relid, RowShareLock);
 577                         erm = (execRowMark *) palloc(sizeof(execRowMark));
 578                         erm->relation = relation;
 579                         erm->rti = rti;
 580                         snprintf(erm->resname, sizeof(erm->resname), "ctid%u", rti);
 581                         estate->es_rowMarks = lappend(estate->es_rowMarks, erm);
 582                 }
 583         }
 584
 585         /*
 586          * initialize the executor "tuple" table.  We need slots for all the
 587          * plan nodes, plus possibly output slots for the junkfilter(s). At
 588          * this point we aren't sure if we need junkfilters, so just add slots
 589          * for them unconditionally.
 590          */
 591         {
 592                 int                     nSlots = ExecCountSlotsNode(plan);
 593
 594                 if (parseTree->resultRelations != NIL)
 595                         nSlots += list_length(parseTree->resultRelations);
 596                 else
 597                         nSlots += 1;
 598                 estate->es_tupleTable = ExecCreateTupleTable(nSlots);
 599         }
 600
 601         /* mark EvalPlanQual not active */
 602         estate->es_topPlan = plan;
 603         estate->es_evalPlanQual = NULL;
 604         estate->es_evTupleNull = NULL;
 605         estate->es_evTuple = NULL;
 606         estate->es_useEvalPlan = false;
 607
 608         /*
 609          * initialize the private state information for all the nodes in the
 610          * query tree.  This opens files, allocates storage and leaves us
 611          * ready to start processing tuples.
 612          */
 613         planstate = ExecInitNode(plan, estate);
 614
 615         /*
 616          * Get the tuple descriptor describing the type of tuples to return.
 617          * (this is especially important if we are creating a relation with
 618          * "SELECT INTO")
 619          */
 620         tupType = ExecGetResultType(planstate);
 621
 622         /*
 623          * Initialize the junk filter if needed.  SELECT and INSERT queries
 624          * need a filter if there are any junk attrs in the tlist.      INSERT and
 625          * SELECT INTO also need a filter if the plan may return raw disk
 626          * tuples (else heap_insert will be scribbling on the source
 627          * relation!). UPDATE and DELETE always need a filter, since there's
 628          * always a junk 'ctid' attribute present --- no need to look first.
 629          */
 630         {
 631                 bool            junk_filter_needed = false;
 632                 ListCell   *tlist;
 633
 634                 switch (operation)
 635                 {
 636                         case CMD_SELECT:
 637                         case CMD_INSERT:
 638                                 foreach(tlist, plan->targetlist)
 639                                 {
 640                                         TargetEntry *tle = (TargetEntry *) lfirst(tlist);
 641
 642                                         if (tle->resjunk)
 643                                         {
 644                                                 junk_filter_needed = true;
 645                                                 break;
 646                                         }
 647                                 }
 648                                 if (!junk_filter_needed &&
 649                                         (operation == CMD_INSERT || do_select_into) &&
 650                                         ExecMayReturnRawTuples(planstate))
 651                                         junk_filter_needed = true;
 652                                 break;
 653                         case CMD_UPDATE:
 654                         case CMD_DELETE:
 655                                 junk_filter_needed = true;
 656                                 break;
 657                         default:
 658                                 break;
 659                 }
 660
 661                 if (junk_filter_needed)
 662                 {
 663                         /*
 664                          * If there are multiple result relations, each one needs its
 665                          * own junk filter.  Note this is only possible for
 666                          * UPDATE/DELETE, so we can't be fooled by some needing a
 667                          * filter and some not.
 668                          */
 669                         if (parseTree->resultRelations != NIL)
 670                         {
 671                                 PlanState **appendplans;
 672                                 int                     as_nplans;
 673                                 ResultRelInfo *resultRelInfo;
 674                                 int                     i;
 675
 676                                 /* Top plan had better be an Append here. */
 677                                 Assert(IsA(plan, Append));
 678                                 Assert(((Append *) plan)->isTarget);
 679                                 Assert(IsA(planstate, AppendState));
 680                                 appendplans = ((AppendState *) planstate)->appendplans;
 681                                 as_nplans = ((AppendState *) planstate)->as_nplans;
 682                                 Assert(as_nplans == estate->es_num_result_relations);
 683                                 resultRelInfo = estate->es_result_relations;
 684                                 for (i = 0; i < as_nplans; i++)
 685                                 {
 686                                         PlanState  *subplan = appendplans[i];
 687                                         JunkFilter *j;
 688
 689                                         j = ExecInitJunkFilter(subplan->plan->targetlist,
 690                                                                                    resultRelInfo->ri_RelationDesc->rd_att->tdhasoid,
 691                                                                                    ExecAllocTableSlot(estate->es_tupleTable));
 692                                         resultRelInfo->ri_junkFilter = j;
 693                                         resultRelInfo++;
 694                                 }
 695
 696                                 /*
 697                                  * Set active junkfilter too; at this point ExecInitAppend
 698                                  * has already selected an active result relation...
 699                                  */
 700                                 estate->es_junkFilter =
 701                                         estate->es_result_relation_info->ri_junkFilter;
 702                         }
 703                         else
 704                         {
 705                                 /* Normal case with just one JunkFilter */
 706                                 JunkFilter *j;
 707
 708                                 j = ExecInitJunkFilter(planstate->plan->targetlist,
 709                                                                            tupType->tdhasoid,
 710                                                           ExecAllocTableSlot(estate->es_tupleTable));
 711                                 estate->es_junkFilter = j;
 712                                 if (estate->es_result_relation_info)
 713                                         estate->es_result_relation_info->ri_junkFilter = j;
 714
 715                                 /* For SELECT, want to return the cleaned tuple type */
 716                                 if (operation == CMD_SELECT)
 717                                         tupType = j->jf_cleanTupType;
 718                         }
 719                 }
 720                 else
 721                         estate->es_junkFilter = NULL;
 722         }
 723
 724         /*
 725          * If doing SELECT INTO, initialize the "into" relation.  We must wait
 726          * till now so we have the "clean" result tuple type to create the new
 727          * table from.
 728          *
 729          * If EXPLAIN, skip creating the "into" relation.
 730          */
 731         intoRelationDesc = NULL;
 732
 733         if (do_select_into && !explainOnly)
 734         {
 735                 char       *intoName;
 736                 Oid                     namespaceId;
 737                 AclResult       aclresult;
 738                 Oid                     intoRelationId;
 739                 TupleDesc       tupdesc;
 740
 741                 /*
 742                  * find namespace to create in, check permissions
 743                  */
 744                 intoName = parseTree->into->relname;
 745                 namespaceId = RangeVarGetCreationNamespace(parseTree->into);
 746
 747                 aclresult = pg_namespace_aclcheck(namespaceId, GetUserId(),
 748                                                                                   ACL_CREATE);
 749                 if (aclresult != ACLCHECK_OK)
 750                         aclcheck_error(aclresult, ACL_KIND_NAMESPACE,
 751                                                    get_namespace_name(namespaceId));
 752
 753                 /*
 754                  * have to copy tupType to get rid of constraints
 755                  */
 756                 tupdesc = CreateTupleDescCopy(tupType);
 757
 758                 intoRelationId = heap_create_with_catalog(intoName,
 759                                                                                                   namespaceId,
 760                                                                                                   InvalidOid,
 761                                                                                                   InvalidOid,
 762                                                                                                   GetUserId(),
 763                                                                                                   tupdesc,
 764                                                                                                   RELKIND_RELATION,
 765                                                                                                   false,
 766                                                                                                   true,
 767                                                                                                   0,
 768                                                                                                   ONCOMMIT_NOOP,
 769                                                                                                   allowSystemTableMods);
 770
 771                 FreeTupleDesc(tupdesc);
 772
 773                 /*
 774                  * Advance command counter so that the newly-created relation's
 775                  * catalog tuples will be visible to heap_open.
 776                  */
 777                 CommandCounterIncrement();
 778
 779                 /*
 780                  * If necessary, create a TOAST table for the into relation. Note
 781                  * that AlterTableCreateToastTable ends with
 782                  * CommandCounterIncrement(), so that the TOAST table will be
 783                  * visible for insertion.
 784                  */
 785                 AlterTableCreateToastTable(intoRelationId, true);
 786
 787                 /*
 788                  * And open the constructed table for writing.
 789                  */
 790                 intoRelationDesc = heap_open(intoRelationId, AccessExclusiveLock);
 791
 792                 /* use_wal off requires rd_targblock be initially invalid */
 793                 Assert(intoRelationDesc->rd_targblock == InvalidBlockNumber);
 794
 795                 /*
 796                  * We can skip WAL-logging the insertions, unless PITR is in use.
 797                  *
 798                  * Note that for a non-temp INTO table, this is safe only because
 799                  * we know that the catalog changes above will have been WAL-logged,
 800                  * and so RecordTransactionCommit will think it needs to WAL-log the
 801                  * eventual transaction commit.  Else the commit might be lost, even
 802                  * though all the data is safely fsync'd ...
 803                  */
 804                 estate->es_into_relation_use_wal = XLogArchivingActive();
 805         }
 806
 807         estate->es_into_relation_descriptor = intoRelationDesc;
 808
 809         queryDesc->tupDesc = tupType;
 810         queryDesc->planstate = planstate;
 811 }
 812
 813 /*
 814  * Initialize ResultRelInfo data for one result relation
 815  */
 816 static void
 817 initResultRelInfo(ResultRelInfo *resultRelInfo,
 818                                   Index resultRelationIndex,
 819                                   List *rangeTable,
 820                                   CmdType operation,
 821                                   bool doInstrument)
 822 {
 823         Oid                     resultRelationOid;
 824         Relation        resultRelationDesc;
 825
 826         resultRelationOid = getrelid(resultRelationIndex, rangeTable);
 827         resultRelationDesc = heap_open(resultRelationOid, RowExclusiveLock);
 828
 829         switch (resultRelationDesc->rd_rel->relkind)
 830         {
 831                 case RELKIND_SEQUENCE:
 832                         ereport(ERROR,
 833                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 834                                          errmsg("cannot change sequence \"%s\"",
 835                                                   RelationGetRelationName(resultRelationDesc))));
 836                         break;
 837                 case RELKIND_TOASTVALUE:
 838                         ereport(ERROR,
 839                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 840                                          errmsg("cannot change TOAST relation \"%s\"",
 841                                                   RelationGetRelationName(resultRelationDesc))));
 842                         break;
 843                 case RELKIND_VIEW:
 844                         ereport(ERROR,
 845                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 846                                          errmsg("cannot change view \"%s\"",
 847                                                   RelationGetRelationName(resultRelationDesc))));
 848                         break;
 849         }
 850
 851         MemSet(resultRelInfo, 0, sizeof(ResultRelInfo));
 852         resultRelInfo->type = T_ResultRelInfo;
 853         resultRelInfo->ri_RangeTableIndex = resultRelationIndex;
 854         resultRelInfo->ri_RelationDesc = resultRelationDesc;
 855         resultRelInfo->ri_NumIndices = 0;
 856         resultRelInfo->ri_IndexRelationDescs = NULL;
 857         resultRelInfo->ri_IndexRelationInfo = NULL;
 858         /* make a copy so as not to depend on relcache info not changing... */
 859         resultRelInfo->ri_TrigDesc = CopyTriggerDesc(resultRelationDesc->trigdesc);
 860         if (resultRelInfo->ri_TrigDesc)
 861         {
 862                 int             n = resultRelInfo->ri_TrigDesc->numtriggers;
 863
 864                 resultRelInfo->ri_TrigFunctions = (FmgrInfo *)
 865                         palloc0(n * sizeof(FmgrInfo));
 866                 if (doInstrument)
 867                         resultRelInfo->ri_TrigInstrument = InstrAlloc(n);
 868                 else
 869                         resultRelInfo->ri_TrigInstrument = NULL;
 870         }
 871         else
 872         {
 873                 resultRelInfo->ri_TrigFunctions = NULL;
 874                 resultRelInfo->ri_TrigInstrument = NULL;
 875         }
 876         resultRelInfo->ri_ConstraintExprs = NULL;
 877         resultRelInfo->ri_junkFilter = NULL;
 878
 879         /*
 880          * If there are indices on the result relation, open them and save
 881          * descriptors in the result relation info, so that we can add new
 882          * index entries for the tuples we add/update.  We need not do this
 883          * for a DELETE, however, since deletion doesn't affect indexes.
 884          */
 885         if (resultRelationDesc->rd_rel->relhasindex &&
 886                 operation != CMD_DELETE)
 887                 ExecOpenIndices(resultRelInfo);
 888 }
 889
 890 /*
 891  *              ExecContextForcesOids
 892  *
 893  * This is pretty grotty: when doing INSERT, UPDATE, or SELECT INTO,
 894  * we need to ensure that result tuples have space for an OID iff they are
 895  * going to be stored into a relation that has OIDs.  In other contexts
 896  * we are free to choose whether to leave space for OIDs in result tuples
 897  * (we generally don't want to, but we do if a physical-tlist optimization
 898  * is possible).  This routine checks the plan context and returns TRUE if the
 899  * choice is forced, FALSE if the choice is not forced.  In the TRUE case,
 900  * *hasoids is set to the required value.
 901  *
 902  * One reason this is ugly is that all plan nodes in the plan tree will emit
 903  * tuples with space for an OID, though we really only need the topmost node
 904  * to do so.  However, node types like Sort don't project new tuples but just
 905  * return their inputs, and in those cases the requirement propagates down
 906  * to the input node.  Eventually we might make this code smart enough to
 907  * recognize how far down the requirement really goes, but for now we just
 908  * make all plan nodes do the same thing if the top level forces the choice.
 909  *
 910  * We assume that estate->es_result_relation_info is already set up to
 911  * describe the target relation.  Note that in an UPDATE that spans an
 912  * inheritance tree, some of the target relations may have OIDs and some not.
 913  * We have to make the decisions on a per-relation basis as we initialize
 914  * each of the child plans of the topmost Append plan.
 915  *
 916  * SELECT INTO is even uglier, because we don't have the INTO relation's
 917  * descriptor available when this code runs; we have to look aside at a
 918  * flag set by InitPlan().
 919  */
 920 bool
 921 ExecContextForcesOids(PlanState *planstate, bool *hasoids)
 922 {
 923         if (planstate->state->es_select_into)
 924         {
 925                 *hasoids = planstate->state->es_into_oids;
 926                 return true;
 927         }
 928         else
 929         {
 930                 ResultRelInfo *ri = planstate->state->es_result_relation_info;
 931
 932                 if (ri != NULL)
 933                 {
 934                         Relation        rel = ri->ri_RelationDesc;
 935
 936                         if (rel != NULL)
 937                         {
 938                                 *hasoids = rel->rd_rel->relhasoids;
 939                                 return true;
 940                         }
 941                 }
 942         }
 943
 944         return false;
 945 }
 946
 947 /* ----------------------------------------------------------------
 948  *              ExecEndPlan
 949  *
 950  *              Cleans up the query plan -- closes files and frees up storage
 951  *
 952  * NOTE: we are no longer very worried about freeing storage per se
 953  * in this code; FreeExecutorState should be guaranteed to release all
 954  * memory that needs to be released.  What we are worried about doing
 955  * is closing relations and dropping buffer pins.  Thus, for example,
 956  * tuple tables must be cleared or dropped to ensure pins are released.
 957  * ----------------------------------------------------------------
 958  */
 959 void
 960 ExecEndPlan(PlanState *planstate, EState *estate)
 961 {
 962         ResultRelInfo *resultRelInfo;
 963         int                     i;
 964         ListCell   *l;
 965
 966         /*
 967          * shut down any PlanQual processing we were doing
 968          */
 969         if (estate->es_evalPlanQual != NULL)
 970                 EndEvalPlanQual(estate);
 971
 972         /*
 973          * shut down the node-type-specific query processing
 974          */
 975         ExecEndNode(planstate);
 976
 977         /*
 978          * destroy the executor "tuple" table.
 979          */
 980         ExecDropTupleTable(estate->es_tupleTable, true);
 981         estate->es_tupleTable = NULL;
 982
 983         /*
 984          * close the result relation(s) if any, but hold locks until xact
 985          * commit.
 986          */
 987         resultRelInfo = estate->es_result_relations;
 988         for (i = estate->es_num_result_relations; i > 0; i--)
 989         {
 990                 /* Close indices and then the relation itself */
 991                 ExecCloseIndices(resultRelInfo);
 992                 heap_close(resultRelInfo->ri_RelationDesc, NoLock);
 993                 resultRelInfo++;
 994         }
 995
 996         /*
 997          * close the "into" relation if necessary, again keeping lock
 998          */
 999         if (estate->es_into_relation_descriptor != NULL)
1000         {
1001                 /*
1002                  * If we skipped using WAL, and it's not a temp relation,
1003                  * we must force the relation down to disk before it's
1004                  * safe to commit the transaction.  This requires forcing
1005                  * out any dirty buffers and then doing a forced fsync.
1006                  */
1007                 if (!estate->es_into_relation_use_wal &&
1008                         !estate->es_into_relation_descriptor->rd_istemp)
1009                 {
1010                         FlushRelationBuffers(estate->es_into_relation_descriptor);
1011                         smgrimmedsync(estate->es_into_relation_descriptor->rd_smgr);
1012                 }
1013
1014                 heap_close(estate->es_into_relation_descriptor, NoLock);
1015         }
1016
1017         /*
1018          * close any relations selected FOR UPDATE/FOR SHARE, again keeping locks
1019          */
1020         foreach(l, estate->es_rowMarks)
1021         {
1022                 execRowMark *erm = lfirst(l);
1023
1024                 heap_close(erm->relation, NoLock);
1025         }
1026 }
1027
1028 /* ----------------------------------------------------------------
1029  *              ExecutePlan
1030  *
1031  *              processes the query plan to retrieve 'numberTuples' tuples in the
1032  *              direction specified.
1033  *
1034  *              Retrieves all tuples if numberTuples is 0
1035  *
1036  *              result is either a slot containing the last tuple in the case
1037  *              of a SELECT or NULL otherwise.
1038  *
1039  * Note: the ctid attribute is a 'junk' attribute that is removed before the
1040  * user can see it
1041  * ----------------------------------------------------------------
1042  */
1043 static TupleTableSlot *
1044 ExecutePlan(EState *estate,
1045                         PlanState *planstate,
1046                         CmdType operation,
1047                         long numberTuples,
1048                         ScanDirection direction,
1049                         DestReceiver *dest)
1050 {
1051         JunkFilter *junkfilter;
1052         TupleTableSlot *slot;
1053         ItemPointer tupleid = NULL;
1054         ItemPointerData tuple_ctid;
1055         long            current_tuple_count;
1056         TupleTableSlot *result;
1057
1058         /*
1059          * initialize local variables
1060          */
1061         slot = NULL;
1062         current_tuple_count = 0;
1063         result = NULL;
1064
1065         /*
1066          * Set the direction.
1067          */
1068         estate->es_direction = direction;
1069
1070         /*
1071          * Process BEFORE EACH STATEMENT triggers
1072          */
1073         switch (operation)
1074         {
1075                 case CMD_UPDATE:
1076                         ExecBSUpdateTriggers(estate, estate->es_result_relation_info);
1077                         break;
1078                 case CMD_DELETE:
1079                         ExecBSDeleteTriggers(estate, estate->es_result_relation_info);
1080                         break;
1081                 case CMD_INSERT:
1082                         ExecBSInsertTriggers(estate, estate->es_result_relation_info);
1083                         break;
1084                 default:
1085                         /* do nothing */
1086                         break;
1087         }
1088
1089         /*
1090          * Loop until we've processed the proper number of tuples from the
1091          * plan.
1092          */
1093
1094         for (;;)
1095         {
1096                 /* Reset the per-output-tuple exprcontext */
1097                 ResetPerTupleExprContext(estate);
1098
1099                 /*
1100                  * Execute the plan and obtain a tuple
1101                  */
1102 lnext:  ;
1103                 if (estate->es_useEvalPlan)
1104                 {
1105                         slot = EvalPlanQualNext(estate);
1106                         if (TupIsNull(slot))
1107                                 slot = ExecProcNode(planstate);
1108                 }
1109                 else
1110                         slot = ExecProcNode(planstate);
1111
1112                 /*
1113                  * if the tuple is null, then we assume there is nothing more to
1114                  * process so we just return null...
1115                  */
1116                 if (TupIsNull(slot))
1117                 {
1118                         result = NULL;
1119                         break;
1120                 }
1121
1122                 /*
1123                  * if we have a junk filter, then project a new tuple with the
1124                  * junk removed.
1125                  *
1126                  * Store this new "clean" tuple in the junkfilter's resultSlot.
1127                  * (Formerly, we stored it back over the "dirty" tuple, which is
1128                  * WRONG because that tuple slot has the wrong descriptor.)
1129                  *
1130                  * Also, extract all the junk information we need.
1131                  */
1132                 if ((junkfilter = estate->es_junkFilter) != NULL)
1133                 {
1134                         Datum           datum;
1135                         bool            isNull;
1136
1137                         /*
1138                          * extract the 'ctid' junk attribute.
1139                          */
1140                         if (operation == CMD_UPDATE || operation == CMD_DELETE)
1141                         {
1142                                 if (!ExecGetJunkAttribute(junkfilter,
1143                                                                                   slot,
1144                                                                                   "ctid",
1145                                                                                   &datum,
1146                                                                                   &isNull))
1147                                         elog(ERROR, "could not find junk ctid column");
1148
1149                                 /* shouldn't ever get a null result... */
1150                                 if (isNull)
1151                                         elog(ERROR, "ctid is NULL");
1152
1153                                 tupleid = (ItemPointer) DatumGetPointer(datum);
1154                                 tuple_ctid = *tupleid;  /* make sure we don't free the
1155                                                                                  * ctid!! */
1156                                 tupleid = &tuple_ctid;
1157                         }
1158                         /*
1159                          * Process any FOR UPDATE or FOR SHARE locking requested.
1160                          */
1161                         else if (estate->es_rowMarks != NIL)
1162                         {
1163                                 ListCell   *l;
1164
1165                 lmark:  ;
1166                                 foreach(l, estate->es_rowMarks)
1167                                 {
1168                                         execRowMark *erm = lfirst(l);
1169                                         HeapTupleData tuple;
1170                                         Buffer          buffer;
1171                                         ItemPointerData update_ctid;
1172                                         TransactionId update_xmax;
1173                                         TupleTableSlot *newSlot;
1174                                         LockTupleMode   lockmode;
1175                                         HTSU_Result             test;
1176
1177                                         if (!ExecGetJunkAttribute(junkfilter,
1178                                                                                           slot,
1179                                                                                           erm->resname,
1180                                                                                           &datum,
1181                                                                                           &isNull))
1182                                                 elog(ERROR, "could not find junk \"%s\" column",
1183                                                          erm->resname);
1184
1185                                         /* shouldn't ever get a null result... */
1186                                         if (isNull)
1187                                                 elog(ERROR, "\"%s\" is NULL", erm->resname);
1188
1189                                         tuple.t_self = *((ItemPointer) DatumGetPointer(datum));
1190
1191                                         if (estate->es_forUpdate)
1192                                                 lockmode = LockTupleExclusive;
1193                                         else
1194                                                 lockmode = LockTupleShared;
1195
1196                                         test = heap_lock_tuple(erm->relation, &tuple, &buffer,
1197                                                                                    &update_ctid, &update_xmax,
1198                                                                                    estate->es_snapshot->curcid,
1199                                                                                    lockmode, estate->es_rowNoWait);
1200                                         ReleaseBuffer(buffer);
1201                                         switch (test)
1202                                         {
1203                                                 case HeapTupleSelfUpdated:
1204                                                         /* treat it as deleted; do not process */
1205                                                         goto lnext;
1206
1207                                                 case HeapTupleMayBeUpdated:
1208                                                         break;
1209
1210                                                 case HeapTupleUpdated:
1211                                                         if (IsXactIsoLevelSerializable)
1212                                                                 ereport(ERROR,
1213                                                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1214                                                                                  errmsg("could not serialize access due to concurrent update")));
1215                                                         if (!ItemPointerEquals(&update_ctid,
1216                                                                                                    &tuple.t_self))
1217                                                         {
1218                                                                 /* updated, so look at updated version */
1219                                                                 newSlot = EvalPlanQual(estate,
1220                                                                                                            erm->rti,
1221                                                                                                            &update_ctid,
1222                                                                                                            update_xmax);
1223                                                                 if (!TupIsNull(newSlot))
1224                                                                 {
1225                                                                         slot = newSlot;
1226                                                                         estate->es_useEvalPlan = true;
1227                                                                         goto lmark;
1228                                                                 }
1229                                                         }
1230
1231                                                         /*
1232                                                          * if tuple was deleted or PlanQual failed for
1233                                                          * updated tuple - we must not return this
1234                                                          * tuple!
1235                                                          */
1236                                                         goto lnext;
1237
1238                                                 default:
1239                                                         elog(ERROR, "unrecognized heap_lock_tuple status: %u",
1240                                                                  test);
1241                                                         return (NULL);
1242                                         }
1243                                 }
1244                         }
1245
1246                         /*
1247                          * Finally create a new "clean" tuple with all junk attributes
1248                          * removed
1249                          */
1250                         slot = ExecFilterJunk(junkfilter, slot);
1251                 }
1252
1253                 /*
1254                  * now that we have a tuple, do the appropriate thing with it..
1255                  * either return it to the user, add it to a relation someplace,
1256                  * delete it from a relation, or modify some of its attributes.
1257                  */
1258                 switch (operation)
1259                 {
1260                         case CMD_SELECT:
1261                                 ExecSelect(slot,        /* slot containing tuple */
1262                                                    dest,        /* destination's tuple-receiver obj */
1263                                                    estate);
1264                                 result = slot;
1265                                 break;
1266
1267                         case CMD_INSERT:
1268                                 ExecInsert(slot, tupleid, estate);
1269                                 result = NULL;
1270                                 break;
1271
1272                         case CMD_DELETE:
1273                                 ExecDelete(slot, tupleid, estate);
1274                                 result = NULL;
1275                                 break;
1276
1277                         case CMD_UPDATE:
1278                                 ExecUpdate(slot, tupleid, estate);
1279                                 result = NULL;
1280                                 break;
1281
1282                         default:
1283                                 elog(ERROR, "unrecognized operation code: %d",
1284                                          (int) operation);
1285                                 result = NULL;
1286                                 break;
1287                 }
1288
1289                 /*
1290                  * check our tuple count.. if we've processed the proper number
1291                  * then quit, else loop again and process more tuples.  Zero
1292                  * numberTuples means no limit.
1293                  */
1294                 current_tuple_count++;
1295                 if (numberTuples && numberTuples == current_tuple_count)
1296                         break;
1297         }
1298
1299         /*
1300          * Process AFTER EACH STATEMENT triggers
1301          */
1302         switch (operation)
1303         {
1304                 case CMD_UPDATE:
1305                         ExecASUpdateTriggers(estate, estate->es_result_relation_info);
1306                         break;
1307                 case CMD_DELETE:
1308                         ExecASDeleteTriggers(estate, estate->es_result_relation_info);
1309                         break;
1310                 case CMD_INSERT:
1311                         ExecASInsertTriggers(estate, estate->es_result_relation_info);
1312                         break;
1313                 default:
1314                         /* do nothing */
1315                         break;
1316         }
1317
1318         /*
1319          * here, result is either a slot containing a tuple in the case of a
1320          * SELECT or NULL otherwise.
1321          */
1322         return result;
1323 }
1324
1325 /* ----------------------------------------------------------------
1326  *              ExecSelect
1327  *
1328  *              SELECTs are easy.. we just pass the tuple to the appropriate
1329  *              print function.  The only complexity is when we do a
1330  *              "SELECT INTO", in which case we insert the tuple into
1331  *              the appropriate relation (note: this is a newly created relation
1332  *              so we don't need to worry about indices or locks.)
1333  * ----------------------------------------------------------------
1334  */
1335 static void
1336 ExecSelect(TupleTableSlot *slot,
1337                    DestReceiver *dest,
1338                    EState *estate)
1339 {
1340         /*
1341          * insert the tuple into the "into relation"
1342          *
1343          * XXX this probably ought to be replaced by a separate destination
1344          */
1345         if (estate->es_into_relation_descriptor != NULL)
1346         {
1347                 HeapTuple       tuple;
1348
1349                 tuple = ExecCopySlotTuple(slot);
1350                 heap_insert(estate->es_into_relation_descriptor, tuple,
1351                                         estate->es_snapshot->curcid,
1352                                         estate->es_into_relation_use_wal,
1353                                         false);         /* never any point in using FSM */
1354                 /* we know there are no indexes to update */
1355                 heap_freetuple(tuple);
1356                 IncrAppended();
1357         }
1358
1359         /*
1360          * send the tuple to the destination
1361          */
1362         (*dest->receiveSlot) (slot, dest);
1363         IncrRetrieved();
1364         (estate->es_processed)++;
1365 }
1366
1367 /* ----------------------------------------------------------------
1368  *              ExecInsert
1369  *
1370  *              INSERTs are trickier.. we have to insert the tuple into
1371  *              the base relation and insert appropriate tuples into the
1372  *              index relations.
1373  * ----------------------------------------------------------------
1374  */
1375 static void
1376 ExecInsert(TupleTableSlot *slot,
1377                    ItemPointer tupleid,
1378                    EState *estate)
1379 {
1380         HeapTuple       tuple;
1381         ResultRelInfo *resultRelInfo;
1382         Relation        resultRelationDesc;
1383         Oid                     newId;
1384
1385         /*
1386          * get the heap tuple out of the tuple table slot, making sure
1387          * we have a writable copy
1388          */
1389         tuple = ExecMaterializeSlot(slot);
1390
1391         /*
1392          * get information on the (current) result relation
1393          */
1394         resultRelInfo = estate->es_result_relation_info;
1395         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1396
1397         /* BEFORE ROW INSERT Triggers */
1398         if (resultRelInfo->ri_TrigDesc &&
1399           resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_INSERT] > 0)
1400         {
1401                 HeapTuple       newtuple;
1402
1403                 newtuple = ExecBRInsertTriggers(estate, resultRelInfo, tuple);
1404
1405                 if (newtuple == NULL)   /* "do nothing" */
1406                         return;
1407
1408                 if (newtuple != tuple)  /* modified by Trigger(s) */
1409                 {
1410                         /*
1411                          * Insert modified tuple into tuple table slot, replacing the
1412                          * original.  We assume that it was allocated in per-tuple
1413                          * memory context, and therefore will go away by itself. The
1414                          * tuple table slot should not try to clear it.
1415                          */
1416                         ExecStoreTuple(newtuple, slot, InvalidBuffer, false);
1417                         tuple = newtuple;
1418                 }
1419         }
1420
1421         /*
1422          * Check the constraints of the tuple
1423          */
1424         if (resultRelationDesc->rd_att->constr)
1425                 ExecConstraints(resultRelInfo, slot, estate);
1426
1427         /*
1428          * insert the tuple
1429          *
1430          * Note: heap_insert returns the tid (location) of the new tuple
1431          * in the t_self field.
1432          */
1433         newId = heap_insert(resultRelationDesc, tuple,
1434                                                 estate->es_snapshot->curcid,
1435                                                 true, true);
1436
1437         IncrAppended();
1438         (estate->es_processed)++;
1439         estate->es_lastoid = newId;
1440         setLastTid(&(tuple->t_self));
1441
1442         /*
1443          * insert index entries for tuple
1444          */
1445         if (resultRelInfo->ri_NumIndices > 0)
1446                 ExecInsertIndexTuples(slot, &(tuple->t_self), estate, false);
1447
1448         /* AFTER ROW INSERT Triggers */
1449         ExecARInsertTriggers(estate, resultRelInfo, tuple);
1450 }
1451
1452 /* ----------------------------------------------------------------
1453  *              ExecDelete
1454  *
1455  *              DELETE is like UPDATE, we delete the tuple and its
1456  *              index tuples.
1457  * ----------------------------------------------------------------
1458  */
1459 static void
1460 ExecDelete(TupleTableSlot *slot,
1461                    ItemPointer tupleid,
1462                    EState *estate)
1463 {
1464         ResultRelInfo *resultRelInfo;
1465         Relation        resultRelationDesc;
1466         HTSU_Result     result;
1467         ItemPointerData update_ctid;
1468         TransactionId update_xmax;
1469
1470         /*
1471          * get information on the (current) result relation
1472          */
1473         resultRelInfo = estate->es_result_relation_info;
1474         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1475
1476         /* BEFORE ROW DELETE Triggers */
1477         if (resultRelInfo->ri_TrigDesc &&
1478           resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_DELETE] > 0)
1479         {
1480                 bool            dodelete;
1481
1482                 dodelete = ExecBRDeleteTriggers(estate, resultRelInfo, tupleid,
1483                                                                                 estate->es_snapshot->curcid);
1484
1485                 if (!dodelete)                  /* "do nothing" */
1486                         return;
1487         }
1488
1489         /*
1490          * delete the tuple
1491          *
1492          * Note: if es_crosscheck_snapshot isn't InvalidSnapshot, we check that
1493          * the row to be deleted is visible to that snapshot, and throw a can't-
1494          * serialize error if not.  This is a special-case behavior needed for
1495          * referential integrity updates in serializable transactions.
1496          */
1497 ldelete:;
1498         result = heap_delete(resultRelationDesc, tupleid,
1499                                                  &update_ctid, &update_xmax,
1500                                                  estate->es_snapshot->curcid,
1501                                                  estate->es_crosscheck_snapshot,
1502                                                  true /* wait for commit */ );
1503         switch (result)
1504         {
1505                 case HeapTupleSelfUpdated:
1506                         /* already deleted by self; nothing to do */
1507                         return;
1508
1509                 case HeapTupleMayBeUpdated:
1510                         break;
1511
1512                 case HeapTupleUpdated:
1513                         if (IsXactIsoLevelSerializable)
1514                                 ereport(ERROR,
1515                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1516                                                  errmsg("could not serialize access due to concurrent update")));
1517                         else if (!ItemPointerEquals(tupleid, &update_ctid))
1518                         {
1519                                 TupleTableSlot *epqslot;
1520
1521                                 epqslot = EvalPlanQual(estate,
1522                                                                            resultRelInfo->ri_RangeTableIndex,
1523                                                                            &update_ctid,
1524                                                                            update_xmax);
1525                                 if (!TupIsNull(epqslot))
1526                                 {
1527                                         *tupleid = update_ctid;
1528                                         goto ldelete;
1529                                 }
1530                         }
1531                         /* tuple already deleted; nothing to do */
1532                         return;
1533
1534                 default:
1535                         elog(ERROR, "unrecognized heap_delete status: %u", result);
1536                         return;
1537         }
1538
1539         IncrDeleted();
1540         (estate->es_processed)++;
1541
1542         /*
1543          * Note: Normally one would think that we have to delete index tuples
1544          * associated with the heap tuple now..
1545          *
1546          * ... but in POSTGRES, we have no need to do this because the vacuum
1547          * daemon automatically opens an index scan and deletes index tuples
1548          * when it finds deleted heap tuples. -cim 9/27/89
1549          */
1550
1551         /* AFTER ROW DELETE Triggers */
1552         ExecARDeleteTriggers(estate, resultRelInfo, tupleid);
1553 }
1554
1555 /* ----------------------------------------------------------------
1556  *              ExecUpdate
1557  *
1558  *              note: we can't run UPDATE queries with transactions
1559  *              off because UPDATEs are actually INSERTs and our
1560  *              scan will mistakenly loop forever, updating the tuple
1561  *              it just inserted..      This should be fixed but until it
1562  *              is, we don't want to get stuck in an infinite loop
1563  *              which corrupts your database..
1564  * ----------------------------------------------------------------
1565  */
1566 static void
1567 ExecUpdate(TupleTableSlot *slot,
1568                    ItemPointer tupleid,
1569                    EState *estate)
1570 {
1571         HeapTuple       tuple;
1572         ResultRelInfo *resultRelInfo;
1573         Relation        resultRelationDesc;
1574         HTSU_Result     result;
1575         ItemPointerData update_ctid;
1576         TransactionId update_xmax;
1577
1578         /*
1579          * abort the operation if not running transactions
1580          */
1581         if (IsBootstrapProcessingMode())
1582                 elog(ERROR, "cannot UPDATE during bootstrap");
1583
1584         /*
1585          * get the heap tuple out of the tuple table slot, making sure
1586          * we have a writable copy
1587          */
1588         tuple = ExecMaterializeSlot(slot);
1589
1590         /*
1591          * get information on the (current) result relation
1592          */
1593         resultRelInfo = estate->es_result_relation_info;
1594         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1595
1596         /* BEFORE ROW UPDATE Triggers */
1597         if (resultRelInfo->ri_TrigDesc &&
1598           resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_UPDATE] > 0)
1599         {
1600                 HeapTuple       newtuple;
1601
1602                 newtuple = ExecBRUpdateTriggers(estate, resultRelInfo,
1603                                                                                 tupleid, tuple,
1604                                                                                 estate->es_snapshot->curcid);
1605
1606                 if (newtuple == NULL)   /* "do nothing" */
1607                         return;
1608
1609                 if (newtuple != tuple)  /* modified by Trigger(s) */
1610                 {
1611                         /*
1612                          * Insert modified tuple into tuple table slot, replacing the
1613                          * original.  We assume that it was allocated in per-tuple
1614                          * memory context, and therefore will go away by itself. The
1615                          * tuple table slot should not try to clear it.
1616                          */
1617                         ExecStoreTuple(newtuple, slot, InvalidBuffer, false);
1618                         tuple = newtuple;
1619                 }
1620         }
1621
1622         /*
1623          * Check the constraints of the tuple
1624          *
1625          * If we generate a new candidate tuple after EvalPlanQual testing, we
1626          * must loop back here and recheck constraints.  (We don't need to
1627          * redo triggers, however.      If there are any BEFORE triggers then
1628          * trigger.c will have done heap_lock_tuple to lock the correct tuple,
1629          * so there's no need to do them again.)
1630          */
1631 lreplace:;
1632         if (resultRelationDesc->rd_att->constr)
1633                 ExecConstraints(resultRelInfo, slot, estate);
1634
1635         /*
1636          * replace the heap tuple
1637          *
1638          * Note: if es_crosscheck_snapshot isn't InvalidSnapshot, we check that
1639          * the row to be updated is visible to that snapshot, and throw a can't-
1640          * serialize error if not.  This is a special-case behavior needed for
1641          * referential integrity updates in serializable transactions.
1642          */
1643         result = heap_update(resultRelationDesc, tupleid, tuple,
1644                                                  &update_ctid, &update_xmax,
1645                                                  estate->es_snapshot->curcid,
1646                                                  estate->es_crosscheck_snapshot,
1647                                                  true /* wait for commit */ );
1648         switch (result)
1649         {
1650                 case HeapTupleSelfUpdated:
1651                         /* already deleted by self; nothing to do */
1652                         return;
1653
1654                 case HeapTupleMayBeUpdated:
1655                         break;
1656
1657                 case HeapTupleUpdated:
1658                         if (IsXactIsoLevelSerializable)
1659                                 ereport(ERROR,
1660                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1661                                                  errmsg("could not serialize access due to concurrent update")));
1662                         else if (!ItemPointerEquals(tupleid, &update_ctid))
1663                         {
1664                                 TupleTableSlot *epqslot;
1665
1666                                 epqslot = EvalPlanQual(estate,
1667                                                                            resultRelInfo->ri_RangeTableIndex,
1668                                                                            &update_ctid,
1669                                                                            update_xmax);
1670                                 if (!TupIsNull(epqslot))
1671                                 {
1672                                         *tupleid = update_ctid;
1673                                         slot = ExecFilterJunk(estate->es_junkFilter, epqslot);
1674                                         tuple = ExecMaterializeSlot(slot);
1675                                         goto lreplace;
1676                                 }
1677                         }
1678                         /* tuple already deleted; nothing to do */
1679                         return;
1680
1681                 default:
1682                         elog(ERROR, "unrecognized heap_update status: %u", result);
1683                         return;
1684         }
1685
1686         IncrReplaced();
1687         (estate->es_processed)++;
1688
1689         /*
1690          * Note: instead of having to update the old index tuples associated
1691          * with the heap tuple, all we do is form and insert new index tuples.
1692          * This is because UPDATEs are actually DELETEs and INSERTs, and index
1693          * tuple deletion is done automagically by the vacuum daemon. All we
1694          * do is insert new index tuples.  -cim 9/27/89
1695          */
1696
1697         /*
1698          * insert index entries for tuple
1699          *
1700          * Note: heap_update returns the tid (location) of the new tuple
1701          * in the t_self field.
1702          */
1703         if (resultRelInfo->ri_NumIndices > 0)
1704                 ExecInsertIndexTuples(slot, &(tuple->t_self), estate, false);
1705
1706         /* AFTER ROW UPDATE Triggers */
1707         ExecARUpdateTriggers(estate, resultRelInfo, tupleid, tuple);
1708 }
1709
1710 static const char *
1711 ExecRelCheck(ResultRelInfo *resultRelInfo,
1712                          TupleTableSlot *slot, EState *estate)
1713 {
1714         Relation        rel = resultRelInfo->ri_RelationDesc;
1715         int                     ncheck = rel->rd_att->constr->num_check;
1716         ConstrCheck *check = rel->rd_att->constr->check;
1717         ExprContext *econtext;
1718         MemoryContext oldContext;
1719         List       *qual;
1720         int                     i;
1721
1722         /*
1723          * If first time through for this result relation, build expression
1724          * nodetrees for rel's constraint expressions.  Keep them in the
1725          * per-query memory context so they'll survive throughout the query.
1726          */
1727         if (resultRelInfo->ri_ConstraintExprs == NULL)
1728         {
1729                 oldContext = MemoryContextSwitchTo(estate->es_query_cxt);
1730                 resultRelInfo->ri_ConstraintExprs =
1731                         (List **) palloc(ncheck * sizeof(List *));
1732                 for (i = 0; i < ncheck; i++)
1733                 {
1734                         /* ExecQual wants implicit-AND form */
1735                         qual = make_ands_implicit(stringToNode(check[i].ccbin));
1736                         resultRelInfo->ri_ConstraintExprs[i] = (List *)
1737                                 ExecPrepareExpr((Expr *) qual, estate);
1738                 }
1739                 MemoryContextSwitchTo(oldContext);
1740         }
1741
1742         /*
1743          * We will use the EState's per-tuple context for evaluating
1744          * constraint expressions (creating it if it's not already there).
1745          */
1746         econtext = GetPerTupleExprContext(estate);
1747
1748         /* Arrange for econtext's scan tuple to be the tuple under test */
1749         econtext->ecxt_scantuple = slot;
1750
1751         /* And evaluate the constraints */
1752         for (i = 0; i < ncheck; i++)
1753         {
1754                 qual = resultRelInfo->ri_ConstraintExprs[i];
1755
1756                 /*
1757                  * NOTE: SQL92 specifies that a NULL result from a constraint
1758                  * expression is not to be treated as a failure.  Therefore, tell
1759                  * ExecQual to return TRUE for NULL.
1760                  */
1761                 if (!ExecQual(qual, econtext, true))
1762                         return check[i].ccname;
1763         }
1764
1765         /* NULL result means no error */
1766         return NULL;
1767 }
1768
1769 void
1770 ExecConstraints(ResultRelInfo *resultRelInfo,
1771                                 TupleTableSlot *slot, EState *estate)
1772 {
1773         Relation        rel = resultRelInfo->ri_RelationDesc;
1774         TupleConstr *constr = rel->rd_att->constr;
1775
1776         Assert(constr);
1777
1778         if (constr->has_not_null)
1779         {
1780                 int                     natts = rel->rd_att->natts;
1781                 int                     attrChk;
1782
1783                 for (attrChk = 1; attrChk <= natts; attrChk++)
1784                 {
1785                         if (rel->rd_att->attrs[attrChk - 1]->attnotnull &&
1786                                 slot_attisnull(slot, attrChk))
1787                                 ereport(ERROR,
1788                                                 (errcode(ERRCODE_NOT_NULL_VIOLATION),
1789                                                  errmsg("null value in column \"%s\" violates not-null constraint",
1790                                         NameStr(rel->rd_att->attrs[attrChk - 1]->attname))));
1791                 }
1792         }
1793
1794         if (constr->num_check > 0)
1795         {
1796                 const char *failed;
1797
1798                 if ((failed = ExecRelCheck(resultRelInfo, slot, estate)) != NULL)
1799                         ereport(ERROR,
1800                                         (errcode(ERRCODE_CHECK_VIOLATION),
1801                                          errmsg("new row for relation \"%s\" violates check constraint \"%s\"",
1802                                                         RelationGetRelationName(rel), failed)));
1803         }
1804 }
1805
1806 /*
1807  * Check a modified tuple to see if we want to process its updated version
1808  * under READ COMMITTED rules.
1809  *
1810  * See backend/executor/README for some info about how this works.
1811  *
1812  *      estate - executor state data
1813  *      rti - rangetable index of table containing tuple
1814  *      *tid - t_ctid from the outdated tuple (ie, next updated version)
1815  *      priorXmax - t_xmax from the outdated tuple
1816  *
1817  * *tid is also an output parameter: it's modified to hold the TID of the
1818  * latest version of the tuple (note this may be changed even on failure)
1819  *
1820  * Returns a slot containing the new candidate update/delete tuple, or
1821  * NULL if we determine we shouldn't process the row.
1822  */
1823 TupleTableSlot *
1824 EvalPlanQual(EState *estate, Index rti,
1825                          ItemPointer tid, TransactionId priorXmax)
1826 {
1827         evalPlanQual *epq;
1828         EState     *epqstate;
1829         Relation        relation;
1830         HeapTupleData tuple;
1831         HeapTuple       copyTuple = NULL;
1832         bool            endNode;
1833
1834         Assert(rti != 0);
1835
1836         /*
1837          * find relation containing target tuple
1838          */
1839         if (estate->es_result_relation_info != NULL &&
1840                 estate->es_result_relation_info->ri_RangeTableIndex == rti)
1841                 relation = estate->es_result_relation_info->ri_RelationDesc;
1842         else
1843         {
1844                 ListCell   *l;
1845
1846                 relation = NULL;
1847                 foreach(l, estate->es_rowMarks)
1848                 {
1849                         if (((execRowMark *) lfirst(l))->rti == rti)
1850                         {
1851                                 relation = ((execRowMark *) lfirst(l))->relation;
1852                                 break;
1853                         }
1854                 }
1855                 if (relation == NULL)
1856                         elog(ERROR, "could not find RowMark for RT index %u", rti);
1857         }
1858
1859         /*
1860          * fetch tid tuple
1861          *
1862          * Loop here to deal with updated or busy tuples
1863          */
1864         tuple.t_self = *tid;
1865         for (;;)
1866         {
1867                 Buffer          buffer;
1868
1869                 if (heap_fetch(relation, SnapshotDirty, &tuple, &buffer, true, NULL))
1870                 {
1871                         /*
1872                          * If xmin isn't what we're expecting, the slot must have been
1873                          * recycled and reused for an unrelated tuple.  This implies
1874                          * that the latest version of the row was deleted, so we need
1875                          * do nothing.  (Should be safe to examine xmin without getting
1876                          * buffer's content lock, since xmin never changes in an existing
1877                          * tuple.)
1878                          */
1879                         if (!TransactionIdEquals(HeapTupleHeaderGetXmin(tuple.t_data),
1880                                                                          priorXmax))
1881                         {
1882                                 ReleaseBuffer(buffer);
1883                                 return NULL;
1884                         }
1885
1886                         /* otherwise xmin should not be dirty... */
1887                         if (TransactionIdIsValid(SnapshotDirty->xmin))
1888                                 elog(ERROR, "t_xmin is uncommitted in tuple to be updated");
1889
1890                         /*
1891                          * If tuple is being updated by other transaction then we have
1892                          * to wait for its commit/abort.
1893                          */
1894                         if (TransactionIdIsValid(SnapshotDirty->xmax))
1895                         {
1896                                 ReleaseBuffer(buffer);
1897                                 XactLockTableWait(SnapshotDirty->xmax);
1898                                 continue;               /* loop back to repeat heap_fetch */
1899                         }
1900
1901                         /*
1902                          * We got tuple - now copy it for use by recheck query.
1903                          */
1904                         copyTuple = heap_copytuple(&tuple);
1905                         ReleaseBuffer(buffer);
1906                         break;
1907                 }
1908
1909                 /*
1910                  * If the referenced slot was actually empty, the latest version
1911                  * of the row must have been deleted, so we need do nothing.
1912                  */
1913                 if (tuple.t_data == NULL)
1914                 {
1915                         ReleaseBuffer(buffer);
1916                         return NULL;
1917                 }
1918
1919                 /*
1920                  * As above, if xmin isn't what we're expecting, do nothing.
1921                  */
1922                 if (!TransactionIdEquals(HeapTupleHeaderGetXmin(tuple.t_data),
1923                                                                  priorXmax))
1924                 {
1925                         ReleaseBuffer(buffer);
1926                         return NULL;
1927                 }
1928
1929                 /*
1930                  * If we get here, the tuple was found but failed SnapshotDirty.
1931                  * Assuming the xmin is either a committed xact or our own xact
1932                  * (as it certainly should be if we're trying to modify the tuple),
1933                  * this must mean that the row was updated or deleted by either
1934                  * a committed xact or our own xact.  If it was deleted, we can
1935                  * ignore it; if it was updated then chain up to the next version
1936                  * and repeat the whole test.
1937                  *
1938                  * As above, it should be safe to examine xmax and t_ctid without
1939                  * the buffer content lock, because they can't be changing.
1940                  */
1941                 if (ItemPointerEquals(&tuple.t_self, &tuple.t_data->t_ctid))
1942                 {
1943                         /* deleted, so forget about it */
1944                         ReleaseBuffer(buffer);
1945                         return NULL;
1946                 }
1947
1948                 /* updated, so look at the updated row */
1949                 tuple.t_self = tuple.t_data->t_ctid;
1950                 /* updated row should have xmin matching this xmax */
1951                 priorXmax = HeapTupleHeaderGetXmax(tuple.t_data);
1952                 ReleaseBuffer(buffer);
1953                 /* loop back to fetch next in chain */
1954         }
1955
1956         /*
1957          * For UPDATE/DELETE we have to return tid of actual row we're
1958          * executing PQ for.
1959          */
1960         *tid = tuple.t_self;
1961
1962         /*
1963          * Need to run a recheck subquery.      Find or create a PQ stack entry.
1964          */
1965         epq = estate->es_evalPlanQual;
1966         endNode = true;
1967
1968         if (epq != NULL && epq->rti == 0)
1969         {
1970                 /* Top PQ stack entry is idle, so re-use it */
1971                 Assert(!(estate->es_useEvalPlan) && epq->next == NULL);
1972                 epq->rti = rti;
1973                 endNode = false;
1974         }
1975
1976         /*
1977          * If this is request for another RTE - Ra, - then we have to check
1978          * wasn't PlanQual requested for Ra already and if so then Ra' row was
1979          * updated again and we have to re-start old execution for Ra and
1980          * forget all what we done after Ra was suspended. Cool? -:))
1981          */
1982         if (epq != NULL && epq->rti != rti &&
1983                 epq->estate->es_evTuple[rti - 1] != NULL)
1984         {
1985                 do
1986                 {
1987                         evalPlanQual *oldepq;
1988
1989                         /* stop execution */
1990                         EvalPlanQualStop(epq);
1991                         /* pop previous PlanQual from the stack */
1992                         oldepq = epq->next;
1993                         Assert(oldepq && oldepq->rti != 0);
1994                         /* push current PQ to freePQ stack */
1995                         oldepq->free = epq;
1996                         epq = oldepq;
1997                         estate->es_evalPlanQual = epq;
1998                 } while (epq->rti != rti);
1999         }
2000
2001         /*
2002          * If we are requested for another RTE then we have to suspend
2003          * execution of current PlanQual and start execution for new one.
2004          */
2005         if (epq == NULL || epq->rti != rti)
2006         {
2007                 /* try to reuse plan used previously */
2008                 evalPlanQual *newepq = (epq != NULL) ? epq->free : NULL;
2009
2010                 if (newepq == NULL)             /* first call or freePQ stack is empty */
2011                 {
2012                         newepq = (evalPlanQual *) palloc0(sizeof(evalPlanQual));
2013                         newepq->free = NULL;
2014                         newepq->estate = NULL;
2015                         newepq->planstate = NULL;
2016                 }
2017                 else
2018                 {
2019                         /* recycle previously used PlanQual */
2020                         Assert(newepq->estate == NULL);
2021                         epq->free = NULL;
2022                 }
2023                 /* push current PQ to the stack */
2024                 newepq->next = epq;
2025                 epq = newepq;
2026                 estate->es_evalPlanQual = epq;
2027                 epq->rti = rti;
2028                 endNode = false;
2029         }
2030
2031         Assert(epq->rti == rti);
2032
2033         /*
2034          * Ok - we're requested for the same RTE.  Unfortunately we still have
2035          * to end and restart execution of the plan, because ExecReScan
2036          * wouldn't ensure that upper plan nodes would reset themselves.  We
2037          * could make that work if insertion of the target tuple were
2038          * integrated with the Param mechanism somehow, so that the upper plan
2039          * nodes know that their children's outputs have changed.
2040          *
2041          * Note that the stack of free evalPlanQual nodes is quite useless at the
2042          * moment, since it only saves us from pallocing/releasing the
2043          * evalPlanQual nodes themselves.  But it will be useful once we
2044          * implement ReScan instead of end/restart for re-using PlanQual
2045          * nodes.
2046          */
2047         if (endNode)
2048         {
2049                 /* stop execution */
2050                 EvalPlanQualStop(epq);
2051         }
2052
2053         /*
2054          * Initialize new recheck query.
2055          *
2056          * Note: if we were re-using PlanQual plans via ExecReScan, we'd need to
2057          * instead copy down changeable state from the top plan (including
2058          * es_result_relation_info, es_junkFilter) and reset locally
2059          * changeable state in the epq (including es_param_exec_vals,
2060          * es_evTupleNull).
2061          */
2062         EvalPlanQualStart(epq, estate, epq->next);
2063
2064         /*
2065          * free old RTE' tuple, if any, and store target tuple where
2066          * relation's scan node will see it
2067          */
2068         epqstate = epq->estate;
2069         if (epqstate->es_evTuple[rti - 1] != NULL)
2070                 heap_freetuple(epqstate->es_evTuple[rti - 1]);
2071         epqstate->es_evTuple[rti - 1] = copyTuple;
2072
2073         return EvalPlanQualNext(estate);
2074 }
2075
2076 static TupleTableSlot *
2077 EvalPlanQualNext(EState *estate)
2078 {
2079         evalPlanQual *epq = estate->es_evalPlanQual;
2080         MemoryContext oldcontext;
2081         TupleTableSlot *slot;
2082
2083         Assert(epq->rti != 0);
2084
2085 lpqnext:;
2086         oldcontext = MemoryContextSwitchTo(epq->estate->es_query_cxt);
2087         slot = ExecProcNode(epq->planstate);
2088         MemoryContextSwitchTo(oldcontext);
2089
2090         /*
2091          * No more tuples for this PQ. Continue previous one.
2092          */
2093         if (TupIsNull(slot))
2094         {
2095                 evalPlanQual *oldepq;
2096
2097                 /* stop execution */
2098                 EvalPlanQualStop(epq);
2099                 /* pop old PQ from the stack */
2100                 oldepq = epq->next;
2101                 if (oldepq == NULL)
2102                 {
2103                         /* this is the first (oldest) PQ - mark as free */
2104                         epq->rti = 0;
2105                         estate->es_useEvalPlan = false;
2106                         /* and continue Query execution */
2107                         return (NULL);
2108                 }
2109                 Assert(oldepq->rti != 0);
2110                 /* push current PQ to freePQ stack */
2111                 oldepq->free = epq;
2112                 epq = oldepq;
2113                 estate->es_evalPlanQual = epq;
2114                 goto lpqnext;
2115         }
2116
2117         return (slot);
2118 }
2119
2120 static void
2121 EndEvalPlanQual(EState *estate)
2122 {
2123         evalPlanQual *epq = estate->es_evalPlanQual;
2124
2125         if (epq->rti == 0)                      /* plans already shutdowned */
2126         {
2127                 Assert(epq->next == NULL);
2128                 return;
2129         }
2130
2131         for (;;)
2132         {
2133                 evalPlanQual *oldepq;
2134
2135                 /* stop execution */
2136                 EvalPlanQualStop(epq);
2137                 /* pop old PQ from the stack */
2138                 oldepq = epq->next;
2139                 if (oldepq == NULL)
2140                 {
2141                         /* this is the first (oldest) PQ - mark as free */
2142                         epq->rti = 0;
2143                         estate->es_useEvalPlan = false;
2144                         break;
2145                 }
2146                 Assert(oldepq->rti != 0);
2147                 /* push current PQ to freePQ stack */
2148                 oldepq->free = epq;
2149                 epq = oldepq;
2150                 estate->es_evalPlanQual = epq;
2151         }
2152 }
2153
2154 /*
2155  * Start execution of one level of PlanQual.
2156  *
2157  * This is a cut-down version of ExecutorStart(): we copy some state from
2158  * the top-level estate rather than initializing it fresh.
2159  */
2160 static void
2161 EvalPlanQualStart(evalPlanQual *epq, EState *estate, evalPlanQual *priorepq)
2162 {
2163         EState     *epqstate;
2164         int                     rtsize;
2165         MemoryContext oldcontext;
2166
2167         rtsize = list_length(estate->es_range_table);
2168
2169         epq->estate = epqstate = CreateExecutorState();
2170
2171         oldcontext = MemoryContextSwitchTo(epqstate->es_query_cxt);
2172
2173         /*
2174          * The epqstates share the top query's copy of unchanging state such
2175          * as the snapshot, rangetable, result-rel info, and external Param
2176          * info. They need their own copies of local state, including a tuple
2177          * table, es_param_exec_vals, etc.
2178          */
2179         epqstate->es_direction = ForwardScanDirection;
2180         epqstate->es_snapshot = estate->es_snapshot;
2181         epqstate->es_crosscheck_snapshot = estate->es_crosscheck_snapshot;
2182         epqstate->es_range_table = estate->es_range_table;
2183         epqstate->es_result_relations = estate->es_result_relations;
2184         epqstate->es_num_result_relations = estate->es_num_result_relations;
2185         epqstate->es_result_relation_info = estate->es_result_relation_info;
2186         epqstate->es_junkFilter = estate->es_junkFilter;
2187         epqstate->es_into_relation_descriptor = estate->es_into_relation_descriptor;
2188         epqstate->es_into_relation_use_wal = estate->es_into_relation_use_wal;
2189         epqstate->es_param_list_info = estate->es_param_list_info;
2190         if (estate->es_topPlan->nParamExec > 0)
2191                 epqstate->es_param_exec_vals = (ParamExecData *)
2192                         palloc0(estate->es_topPlan->nParamExec * sizeof(ParamExecData));
2193         epqstate->es_rowMarks = estate->es_rowMarks;
2194         epqstate->es_forUpdate = estate->es_forUpdate;
2195         epqstate->es_rowNoWait = estate->es_rowNoWait;
2196         epqstate->es_instrument = estate->es_instrument;
2197         epqstate->es_select_into = estate->es_select_into;
2198         epqstate->es_into_oids = estate->es_into_oids;
2199         epqstate->es_topPlan = estate->es_topPlan;
2200
2201         /*
2202          * Each epqstate must have its own es_evTupleNull state, but all the
2203          * stack entries share es_evTuple state.  This allows sub-rechecks to
2204          * inherit the value being examined by an outer recheck.
2205          */
2206         epqstate->es_evTupleNull = (bool *) palloc0(rtsize * sizeof(bool));
2207         if (priorepq == NULL)
2208                 /* first PQ stack entry */
2209                 epqstate->es_evTuple = (HeapTuple *)
2210                         palloc0(rtsize * sizeof(HeapTuple));
2211         else
2212                 /* later stack entries share the same storage */
2213                 epqstate->es_evTuple = priorepq->estate->es_evTuple;
2214
2215         epqstate->es_tupleTable =
2216                 ExecCreateTupleTable(estate->es_tupleTable->size);
2217
2218         epq->planstate = ExecInitNode(estate->es_topPlan, epqstate);
2219
2220         MemoryContextSwitchTo(oldcontext);
2221 }
2222
2223 /*
2224  * End execution of one level of PlanQual.
2225  *
2226  * This is a cut-down version of ExecutorEnd(); basically we want to do most
2227  * of the normal cleanup, but *not* close result relations (which we are
2228  * just sharing from the outer query).
2229  */
2230 static void
2231 EvalPlanQualStop(evalPlanQual *epq)
2232 {
2233         EState     *epqstate = epq->estate;
2234         MemoryContext oldcontext;
2235
2236         oldcontext = MemoryContextSwitchTo(epqstate->es_query_cxt);
2237
2238         ExecEndNode(epq->planstate);
2239
2240         ExecDropTupleTable(epqstate->es_tupleTable, true);
2241         epqstate->es_tupleTable = NULL;
2242
2243         if (epqstate->es_evTuple[epq->rti - 1] != NULL)
2244         {
2245                 heap_freetuple(epqstate->es_evTuple[epq->rti - 1]);
2246                 epqstate->es_evTuple[epq->rti - 1] = NULL;
2247         }
2248
2249         MemoryContextSwitchTo(oldcontext);
2250
2251         FreeExecutorState(epqstate);
2252
2253         epq->estate = NULL;
2254         epq->planstate = NULL;
2255 }