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  11  *      The old ExecutorMain() has been replaced by ExecutorStart(),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) 1996-2006, 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 1.270 2006/04/30 18:30:38 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/tablespace.h"
  41 #include "commands/trigger.h"
  42 #include "executor/execdebug.h"
  43 #include "executor/execdefs.h"
  44 #include "executor/instrument.h"
  45 #include "miscadmin.h"
  46 #include "optimizer/clauses.h"
  47 #include "optimizer/var.h"
  48 #include "parser/parsetree.h"
  49 #include "storage/smgr.h"
  50 #include "utils/acl.h"
  51 #include "utils/guc.h"
  52 #include "utils/lsyscache.h"
  53 #include "utils/memutils.h"
  54
  55
  56 typedef struct evalPlanQual
  57 {
  58         Index           rti;
  59         EState     *estate;
  60         PlanState  *planstate;
  61         struct evalPlanQual *next;      /* stack of active PlanQual plans */
  62         struct evalPlanQual *free;      /* list of free PlanQual plans */
  63 } evalPlanQual;
  64
  65 /* decls for local routines only used within this module */
  66 static void InitPlan(QueryDesc *queryDesc, int eflags);
  67 static void initResultRelInfo(ResultRelInfo *resultRelInfo,
  68                                   Index resultRelationIndex,
  69                                   List *rangeTable,
  70                                   CmdType operation,
  71                                   bool doInstrument);
  72 static TupleTableSlot *ExecutePlan(EState *estate, PlanState *planstate,
  73                         CmdType operation,
  74                         long numberTuples,
  75                         ScanDirection direction,
  76                         DestReceiver *dest);
  77 static void ExecSelect(TupleTableSlot *slot,
  78                    DestReceiver *dest,
  79                    EState *estate);
  80 static void ExecInsert(TupleTableSlot *slot, ItemPointer tupleid,
  81                    EState *estate);
  82 static void ExecDelete(TupleTableSlot *slot, ItemPointer tupleid,
  83                    EState *estate);
  84 static void ExecUpdate(TupleTableSlot *slot, ItemPointer tupleid,
  85                    EState *estate);
  86 static TupleTableSlot *EvalPlanQualNext(EState *estate);
  87 static void EndEvalPlanQual(EState *estate);
  88 static void ExecCheckRTEPerms(RangeTblEntry *rte);
  89 static void ExecCheckXactReadOnly(Query *parsetree);
  90 static void EvalPlanQualStart(evalPlanQual *epq, EState *estate,
  91                                   evalPlanQual *priorepq);
  92 static void EvalPlanQualStop(evalPlanQual *epq);
  93
  94 /* end of local decls */
  95
  96
  97 /* ----------------------------------------------------------------
  98  *              ExecutorStart
  99  *
 100  *              This routine must be called at the beginning of any execution of any
 101  *              query plan
 102  *
 103  * Takes a QueryDesc previously created by CreateQueryDesc (it's not real
 104  * clear why we bother to separate the two functions, but...).  The tupDesc
 105  * field of the QueryDesc is filled in to describe the tuples that will be
 106  * returned, and the internal fields (estate and planstate) are set up.
 107  *
 108  * eflags contains flag bits as described in executor.h.
 109  *
 110  * NB: the CurrentMemoryContext when this is called will become the parent
 111  * of the per-query context used for this Executor invocation.
 112  * ----------------------------------------------------------------
 113  */
 114 void
 115 ExecutorStart(QueryDesc *queryDesc, int eflags)
 116 {
 117         EState     *estate;
 118         MemoryContext oldcontext;
 119
 120         /* sanity checks: queryDesc must not be started already */
 121         Assert(queryDesc != NULL);
 122         Assert(queryDesc->estate == NULL);
 123
 124         /*
 125          * If the transaction is read-only, we need to check if any writes are
 126          * planned to non-temporary tables.  EXPLAIN is considered read-only.
 127          */
 128         if (XactReadOnly && !(eflags & EXEC_FLAG_EXPLAIN_ONLY))
 129                 ExecCheckXactReadOnly(queryDesc->parsetree);
 130
 131         /*
 132          * Build EState, switch into per-query memory context for startup.
 133          */
 134         estate = CreateExecutorState();
 135         queryDesc->estate = estate;
 136
 137         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 138
 139         /*
 140          * Fill in parameters, if any, from queryDesc
 141          */
 142         estate->es_param_list_info = queryDesc->params;
 143
 144         if (queryDesc->plantree->nParamExec > 0)
 145                 estate->es_param_exec_vals = (ParamExecData *)
 146                         palloc0(queryDesc->plantree->nParamExec * sizeof(ParamExecData));
 147
 148         /*
 149          * Copy other important information into the EState
 150          */
 151         estate->es_snapshot = queryDesc->snapshot;
 152         estate->es_crosscheck_snapshot = queryDesc->crosscheck_snapshot;
 153         estate->es_instrument = queryDesc->doInstrument;
 154
 155         /*
 156          * Initialize the plan state tree
 157          */
 158         InitPlan(queryDesc, eflags);
 159
 160         MemoryContextSwitchTo(oldcontext);
 161 }
 162
 163 /* ----------------------------------------------------------------
 164  *              ExecutorRun
 165  *
 166  *              This is the main routine of the executor module. It accepts
 167  *              the query descriptor from the traffic cop and executes the
 168  *              query plan.
 169  *
 170  *              ExecutorStart must have been called already.
 171  *
 172  *              If direction is NoMovementScanDirection then nothing is done
 173  *              except to start up/shut down the destination.  Otherwise,
 174  *              we retrieve up to 'count' tuples in the specified direction.
 175  *
 176  *              Note: count = 0 is interpreted as no portal limit, i.e., run to
 177  *              completion.
 178  *
 179  * ----------------------------------------------------------------
 180  */
 181 TupleTableSlot *
 182 ExecutorRun(QueryDesc *queryDesc,
 183                         ScanDirection direction, long count)
 184 {
 185         EState     *estate;
 186         CmdType         operation;
 187         DestReceiver *dest;
 188         TupleTableSlot *result;
 189         MemoryContext oldcontext;
 190
 191         /* sanity checks */
 192         Assert(queryDesc != NULL);
 193
 194         estate = queryDesc->estate;
 195
 196         Assert(estate != NULL);
 197
 198         /*
 199          * Switch into per-query memory context
 200          */
 201         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 202
 203         /*
 204          * extract information from the query descriptor and the query feature.
 205          */
 206         operation = queryDesc->operation;
 207         dest = queryDesc->dest;
 208
 209         /*
 210          * startup tuple receiver
 211          */
 212         estate->es_processed = 0;
 213         estate->es_lastoid = InvalidOid;
 214
 215         (*dest->rStartup) (dest, operation, queryDesc->tupDesc);
 216
 217         /*
 218          * run plan
 219          */
 220         if (ScanDirectionIsNoMovement(direction))
 221                 result = NULL;
 222         else
 223                 result = ExecutePlan(estate,
 224                                                          queryDesc->planstate,
 225                                                          operation,
 226                                                          count,
 227                                                          direction,
 228                                                          dest);
 229
 230         /*
 231          * shutdown receiver
 232          */
 233         (*dest->rShutdown) (dest);
 234
 235         MemoryContextSwitchTo(oldcontext);
 236
 237         return result;
 238 }
 239
 240 /* ----------------------------------------------------------------
 241  *              ExecutorEnd
 242  *
 243  *              This routine must be called at the end of execution of any
 244  *              query plan
 245  * ----------------------------------------------------------------
 246  */
 247 void
 248 ExecutorEnd(QueryDesc *queryDesc)
 249 {
 250         EState     *estate;
 251         MemoryContext oldcontext;
 252
 253         /* sanity checks */
 254         Assert(queryDesc != NULL);
 255
 256         estate = queryDesc->estate;
 257
 258         Assert(estate != NULL);
 259
 260         /*
 261          * Switch into per-query memory context to run ExecEndPlan
 262          */
 263         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 264
 265         ExecEndPlan(queryDesc->planstate, estate);
 266
 267         /*
 268          * Must switch out of context before destroying it
 269          */
 270         MemoryContextSwitchTo(oldcontext);
 271
 272         /*
 273          * Release EState and per-query memory context.  This should release
 274          * everything the executor has allocated.
 275          */
 276         FreeExecutorState(estate);
 277
 278         /* Reset queryDesc fields that no longer point to anything */
 279         queryDesc->tupDesc = NULL;
 280         queryDesc->estate = NULL;
 281         queryDesc->planstate = NULL;
 282 }
 283
 284 /* ----------------------------------------------------------------
 285  *              ExecutorRewind
 286  *
 287  *              This routine may be called on an open queryDesc to rewind it
 288  *              to the start.
 289  * ----------------------------------------------------------------
 290  */
 291 void
 292 ExecutorRewind(QueryDesc *queryDesc)
 293 {
 294         EState     *estate;
 295         MemoryContext oldcontext;
 296
 297         /* sanity checks */
 298         Assert(queryDesc != NULL);
 299
 300         estate = queryDesc->estate;
 301
 302         Assert(estate != NULL);
 303
 304         /* It's probably not sensible to rescan updating queries */
 305         Assert(queryDesc->operation == CMD_SELECT);
 306
 307         /*
 308          * Switch into per-query memory context
 309          */
 310         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 311
 312         /*
 313          * rescan plan
 314          */
 315         ExecReScan(queryDesc->planstate, NULL);
 316
 317         MemoryContextSwitchTo(oldcontext);
 318 }
 319
 320
 321 /*
 322  * ExecCheckRTPerms
 323  *              Check access permissions for all relations listed in a range table.
 324  */
 325 void
 326 ExecCheckRTPerms(List *rangeTable)
 327 {
 328         ListCell   *l;
 329
 330         foreach(l, rangeTable)
 331         {
 332                 RangeTblEntry *rte = lfirst(l);
 333
 334                 ExecCheckRTEPerms(rte);
 335         }
 336 }
 337
 338 /*
 339  * ExecCheckRTEPerms
 340  *              Check access permissions for a single RTE.
 341  */
 342 static void
 343 ExecCheckRTEPerms(RangeTblEntry *rte)
 344 {
 345         AclMode         requiredPerms;
 346         Oid                     relOid;
 347         Oid                     userid;
 348
 349         /*
 350          * Only plain-relation RTEs need to be checked here.  Subquery RTEs are
 351          * checked by ExecInitSubqueryScan if the subquery is still a separate
 352          * subquery --- if it's been pulled up into our query level then the RTEs
 353          * are in our rangetable and will be checked here. Function RTEs are
 354          * checked by init_fcache when the function is prepared for execution.
 355          * Join and special RTEs need no checks.
 356          */
 357         if (rte->rtekind != RTE_RELATION)
 358                 return;
 359
 360         /*
 361          * No work if requiredPerms is empty.
 362          */
 363         requiredPerms = rte->requiredPerms;
 364         if (requiredPerms == 0)
 365                 return;
 366
 367         relOid = rte->relid;
 368
 369         /*
 370          * userid to check as: current user unless we have a setuid indication.
 371          *
 372          * Note: GetUserId() is presently fast enough that there's no harm in
 373          * calling it separately for each RTE.  If that stops being true, we could
 374          * call it once in ExecCheckRTPerms and pass the userid down from there.
 375          * But for now, no need for the extra clutter.
 376          */
 377         userid = rte->checkAsUser ? rte->checkAsUser : GetUserId();
 378
 379         /*
 380          * We must have *all* the requiredPerms bits, so use aclmask not aclcheck.
 381          */
 382         if (pg_class_aclmask(relOid, userid, requiredPerms, ACLMASK_ALL)
 383                 != requiredPerms)
 384                 aclcheck_error(ACLCHECK_NO_PRIV, ACL_KIND_CLASS,
 385                                            get_rel_name(relOid));
 386 }
 387
 388 /*
 389  * Check that the query does not imply any writes to non-temp tables.
 390  */
 391 static void
 392 ExecCheckXactReadOnly(Query *parsetree)
 393 {
 394         ListCell   *l;
 395
 396         /*
 397          * CREATE TABLE AS or SELECT INTO?
 398          *
 399          * XXX should we allow this if the destination is temp?
 400          */
 401         if (parsetree->into != NULL)
 402                 goto fail;
 403
 404         /* Fail if write permissions are requested on any non-temp table */
 405         foreach(l, parsetree->rtable)
 406         {
 407                 RangeTblEntry *rte = lfirst(l);
 408
 409                 if (rte->rtekind == RTE_SUBQUERY)
 410                 {
 411                         ExecCheckXactReadOnly(rte->subquery);
 412                         continue;
 413                 }
 414
 415                 if (rte->rtekind != RTE_RELATION)
 416                         continue;
 417
 418                 if ((rte->requiredPerms & (~ACL_SELECT)) == 0)
 419                         continue;
 420
 421                 if (isTempNamespace(get_rel_namespace(rte->relid)))
 422                         continue;
 423
 424                 goto fail;
 425         }
 426
 427         return;
 428
 429 fail:
 430         ereport(ERROR,
 431                         (errcode(ERRCODE_READ_ONLY_SQL_TRANSACTION),
 432                          errmsg("transaction is read-only")));
 433 }
 434
 435
 436 /* ----------------------------------------------------------------
 437  *              InitPlan
 438  *
 439  *              Initializes the query plan: open files, allocate storage
 440  *              and start up the rule manager
 441  * ----------------------------------------------------------------
 442  */
 443 static void
 444 InitPlan(QueryDesc *queryDesc, int eflags)
 445 {
 446         CmdType         operation = queryDesc->operation;
 447         Query      *parseTree = queryDesc->parsetree;
 448         Plan       *plan = queryDesc->plantree;
 449         EState     *estate = queryDesc->estate;
 450         PlanState  *planstate;
 451         List       *rangeTable;
 452         Relation        intoRelationDesc;
 453         bool            do_select_into;
 454         TupleDesc       tupType;
 455         ListCell   *l;
 456
 457         /*
 458          * Do permissions checks.  It's sufficient to examine the query's top
 459          * rangetable here --- subplan RTEs will be checked during
 460          * ExecInitSubPlan().
 461          */
 462         ExecCheckRTPerms(parseTree->rtable);
 463
 464         /*
 465          * get information from query descriptor
 466          */
 467         rangeTable = parseTree->rtable;
 468
 469         /*
 470          * initialize the node's execution state
 471          */
 472         estate->es_range_table = rangeTable;
 473
 474         /*
 475          * if there is a result relation, initialize result relation stuff
 476          */
 477         if (parseTree->resultRelation != 0 && operation != CMD_SELECT)
 478         {
 479                 List       *resultRelations = parseTree->resultRelations;
 480                 int                     numResultRelations;
 481                 ResultRelInfo *resultRelInfos;
 482
 483                 if (resultRelations != NIL)
 484                 {
 485                         /*
 486                          * Multiple result relations (due to inheritance)
 487                          * parseTree->resultRelations identifies them all
 488                          */
 489                         ResultRelInfo *resultRelInfo;
 490
 491                         numResultRelations = list_length(resultRelations);
 492                         resultRelInfos = (ResultRelInfo *)
 493                                 palloc(numResultRelations * sizeof(ResultRelInfo));
 494                         resultRelInfo = resultRelInfos;
 495                         foreach(l, resultRelations)
 496                         {
 497                                 initResultRelInfo(resultRelInfo,
 498                                                                   lfirst_int(l),
 499                                                                   rangeTable,
 500                                                                   operation,
 501                                                                   estate->es_instrument);
 502                                 resultRelInfo++;
 503                         }
 504                 }
 505                 else
 506                 {
 507                         /*
 508                          * Single result relation identified by parseTree->resultRelation
 509                          */
 510                         numResultRelations = 1;
 511                         resultRelInfos = (ResultRelInfo *) palloc(sizeof(ResultRelInfo));
 512                         initResultRelInfo(resultRelInfos,
 513                                                           parseTree->resultRelation,
 514                                                           rangeTable,
 515                                                           operation,
 516                                                           estate->es_instrument);
 517                 }
 518
 519                 estate->es_result_relations = resultRelInfos;
 520                 estate->es_num_result_relations = numResultRelations;
 521                 /* Initialize to first or only result rel */
 522                 estate->es_result_relation_info = resultRelInfos;
 523         }
 524         else
 525         {
 526                 /*
 527                  * if no result relation, then set state appropriately
 528                  */
 529                 estate->es_result_relations = NULL;
 530                 estate->es_num_result_relations = 0;
 531                 estate->es_result_relation_info = NULL;
 532         }
 533
 534         /*
 535          * Detect whether we're doing SELECT INTO.  If so, set the es_into_oids
 536          * flag appropriately so that the plan tree will be initialized with the
 537          * correct tuple descriptors.
 538          */
 539         do_select_into = false;
 540
 541         if (operation == CMD_SELECT && parseTree->into != NULL)
 542         {
 543                 do_select_into = true;
 544                 estate->es_select_into = true;
 545                 estate->es_into_oids = parseTree->intoHasOids;
 546         }
 547
 548         /*
 549          * Have to lock relations selected FOR UPDATE/FOR SHARE
 550          */
 551         estate->es_rowMarks = NIL;
 552         foreach(l, parseTree->rowMarks)
 553         {
 554                 RowMarkClause *rc = (RowMarkClause *) lfirst(l);
 555                 Oid                     relid = getrelid(rc->rti, rangeTable);
 556                 Relation        relation;
 557                 ExecRowMark *erm;
 558
 559                 relation = heap_open(relid, RowShareLock);
 560                 erm = (ExecRowMark *) palloc(sizeof(ExecRowMark));
 561                 erm->relation = relation;
 562                 erm->rti = rc->rti;
 563                 erm->forUpdate = rc->forUpdate;
 564                 erm->noWait = rc->noWait;
 565                 snprintf(erm->resname, sizeof(erm->resname), "ctid%u", rc->rti);
 566                 estate->es_rowMarks = lappend(estate->es_rowMarks, erm);
 567         }
 568
 569         /*
 570          * initialize the executor "tuple" table.  We need slots for all the plan
 571          * nodes, plus possibly output slots for the junkfilter(s). At this point
 572          * we aren't sure if we need junkfilters, so just add slots for them
 573          * unconditionally.  Also, if it's not a SELECT, set up a slot for use for
 574          * trigger output tuples.
 575          */
 576         {
 577                 int                     nSlots = ExecCountSlotsNode(plan);
 578
 579                 if (parseTree->resultRelations != NIL)
 580                         nSlots += list_length(parseTree->resultRelations);
 581                 else
 582                         nSlots += 1;
 583                 if (operation != CMD_SELECT)
 584                         nSlots++;
 585
 586                 estate->es_tupleTable = ExecCreateTupleTable(nSlots);
 587
 588                 if (operation != CMD_SELECT)
 589                         estate->es_trig_tuple_slot =
 590                                 ExecAllocTableSlot(estate->es_tupleTable);
 591         }
 592
 593         /* mark EvalPlanQual not active */
 594         estate->es_topPlan = plan;
 595         estate->es_evalPlanQual = NULL;
 596         estate->es_evTupleNull = NULL;
 597         estate->es_evTuple = NULL;
 598         estate->es_useEvalPlan = false;
 599
 600         /*
 601          * initialize the private state information for all the nodes in the query
 602          * tree.  This opens files, allocates storage and leaves us ready to start
 603          * processing tuples.
 604          */
 605         planstate = ExecInitNode(plan, estate, eflags);
 606
 607         /*
 608          * Get the tuple descriptor describing the type of tuples to return. (this
 609          * is especially important if we are creating a relation with "SELECT
 610          * INTO")
 611          */
 612         tupType = ExecGetResultType(planstate);
 613
 614         /*
 615          * Initialize the junk filter if needed.  SELECT and INSERT queries need a
 616          * filter if there are any junk attrs in the tlist.  INSERT and SELECT
 617          * INTO also need a filter if the plan may return raw disk tuples (else
 618          * heap_insert will be scribbling on the source relation!). UPDATE and
 619          * DELETE always need a filter, since there's always a junk 'ctid'
 620          * attribute present --- no need to look first.
 621          */
 622         {
 623                 bool            junk_filter_needed = false;
 624                 ListCell   *tlist;
 625
 626                 switch (operation)
 627                 {
 628                         case CMD_SELECT:
 629                         case CMD_INSERT:
 630                                 foreach(tlist, plan->targetlist)
 631                                 {
 632                                         TargetEntry *tle = (TargetEntry *) lfirst(tlist);
 633
 634                                         if (tle->resjunk)
 635                                         {
 636                                                 junk_filter_needed = true;
 637                                                 break;
 638                                         }
 639                                 }
 640                                 if (!junk_filter_needed &&
 641                                         (operation == CMD_INSERT || do_select_into) &&
 642                                         ExecMayReturnRawTuples(planstate))
 643                                         junk_filter_needed = true;
 644                                 break;
 645                         case CMD_UPDATE:
 646                         case CMD_DELETE:
 647                                 junk_filter_needed = true;
 648                                 break;
 649                         default:
 650                                 break;
 651                 }
 652
 653                 if (junk_filter_needed)
 654                 {
 655                         /*
 656                          * If there are multiple result relations, each one needs its own
 657                          * junk filter.  Note this is only possible for UPDATE/DELETE, so
 658                          * we can't be fooled by some needing a filter and some not.
 659                          */
 660                         if (parseTree->resultRelations != NIL)
 661                         {
 662                                 PlanState **appendplans;
 663                                 int                     as_nplans;
 664                                 ResultRelInfo *resultRelInfo;
 665                                 int                     i;
 666
 667                                 /* Top plan had better be an Append here. */
 668                                 Assert(IsA(plan, Append));
 669                                 Assert(((Append *) plan)->isTarget);
 670                                 Assert(IsA(planstate, AppendState));
 671                                 appendplans = ((AppendState *) planstate)->appendplans;
 672                                 as_nplans = ((AppendState *) planstate)->as_nplans;
 673                                 Assert(as_nplans == estate->es_num_result_relations);
 674                                 resultRelInfo = estate->es_result_relations;
 675                                 for (i = 0; i < as_nplans; i++)
 676                                 {
 677                                         PlanState  *subplan = appendplans[i];
 678                                         JunkFilter *j;
 679
 680                                         j = ExecInitJunkFilter(subplan->plan->targetlist,
 681                                                         resultRelInfo->ri_RelationDesc->rd_att->tdhasoid,
 682                                                                   ExecAllocTableSlot(estate->es_tupleTable));
 683                                         resultRelInfo->ri_junkFilter = j;
 684                                         resultRelInfo++;
 685                                 }
 686
 687                                 /*
 688                                  * Set active junkfilter too; at this point ExecInitAppend has
 689                                  * already selected an active result relation...
 690                                  */
 691                                 estate->es_junkFilter =
 692                                         estate->es_result_relation_info->ri_junkFilter;
 693                         }
 694                         else
 695                         {
 696                                 /* Normal case with just one JunkFilter */
 697                                 JunkFilter *j;
 698
 699                                 j = ExecInitJunkFilter(planstate->plan->targetlist,
 700                                                                            tupType->tdhasoid,
 701                                                                   ExecAllocTableSlot(estate->es_tupleTable));
 702                                 estate->es_junkFilter = j;
 703                                 if (estate->es_result_relation_info)
 704                                         estate->es_result_relation_info->ri_junkFilter = j;
 705
 706                                 /* For SELECT, want to return the cleaned tuple type */
 707                                 if (operation == CMD_SELECT)
 708                                         tupType = j->jf_cleanTupType;
 709                         }
 710                 }
 711                 else
 712                         estate->es_junkFilter = NULL;
 713         }
 714
 715         /*
 716          * If doing SELECT INTO, initialize the "into" relation.  We must wait
 717          * till now so we have the "clean" result tuple type to create the new
 718          * table from.
 719          *
 720          * If EXPLAIN, skip creating the "into" relation.
 721          */
 722         intoRelationDesc = NULL;
 723
 724         if (do_select_into && !(eflags & EXEC_FLAG_EXPLAIN_ONLY))
 725         {
 726                 char       *intoName;
 727                 Oid                     namespaceId;
 728                 Oid                     tablespaceId;
 729                 AclResult       aclresult;
 730                 Oid                     intoRelationId;
 731                 TupleDesc       tupdesc;
 732
 733                 /*
 734                  * Check consistency of arguments
 735                  */
 736                 if (parseTree->intoOnCommit != ONCOMMIT_NOOP && !parseTree->into->istemp)
 737                         ereport(ERROR,
 738                                         (errcode(ERRCODE_INVALID_TABLE_DEFINITION),
 739                                          errmsg("ON COMMIT can only be used on temporary tables")));
 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                  * Select tablespace to use.  If not specified, use default_tablespace
 755                  * (which may in turn default to database's default).
 756                  */
 757                 if (parseTree->intoTableSpaceName)
 758                 {
 759                         tablespaceId = get_tablespace_oid(parseTree->intoTableSpaceName);
 760                         if (!OidIsValid(tablespaceId))
 761                                 ereport(ERROR,
 762                                                 (errcode(ERRCODE_UNDEFINED_OBJECT),
 763                                                  errmsg("tablespace \"%s\" does not exist",
 764                                                                 parseTree->intoTableSpaceName)));
 765                 } else
 766                 {
 767                         tablespaceId = GetDefaultTablespace();
 768                         /* note InvalidOid is OK in this case */
 769                 }
 770
 771                 /* Check permissions except when using the database's default */
 772                 if (OidIsValid(tablespaceId))
 773                 {
 774                         AclResult       aclresult;
 775
 776                         aclresult = pg_tablespace_aclcheck(tablespaceId, GetUserId(),
 777                                                                                            ACL_CREATE);
 778
 779                         if (aclresult != ACLCHECK_OK)
 780                                 aclcheck_error(aclresult, ACL_KIND_TABLESPACE,
 781                                                            get_tablespace_name(tablespaceId));
 782                 }
 783
 784                 /*
 785                  * have to copy tupType to get rid of constraints
 786                  */
 787                 tupdesc = CreateTupleDescCopy(tupType);
 788
 789                 intoRelationId = heap_create_with_catalog(intoName,
 790                                                                                                   namespaceId,
 791                                                                                                   tablespaceId,
 792                                                                                                   InvalidOid,
 793                                                                                                   GetUserId(),
 794                                                                                                   tupdesc,
 795                                                                                                   RELKIND_RELATION,
 796                                                                                                   false,
 797                                                                                                   true,
 798                                                                                                   0,
 799                                                                                                   parseTree->intoOnCommit,
 800                                                                                                   allowSystemTableMods);
 801
 802                 FreeTupleDesc(tupdesc);
 803
 804                 /*
 805                  * Advance command counter so that the newly-created relation's
 806                  * catalog tuples will be visible to heap_open.
 807                  */
 808                 CommandCounterIncrement();
 809
 810                 /*
 811                  * If necessary, create a TOAST table for the into relation. Note that
 812                  * AlterTableCreateToastTable ends with CommandCounterIncrement(), so
 813                  * that the TOAST table will be visible for insertion.
 814                  */
 815                 AlterTableCreateToastTable(intoRelationId, true);
 816
 817                 /*
 818                  * And open the constructed table for writing.
 819                  */
 820                 intoRelationDesc = heap_open(intoRelationId, AccessExclusiveLock);
 821
 822                 /* use_wal off requires rd_targblock be initially invalid */
 823                 Assert(intoRelationDesc->rd_targblock == InvalidBlockNumber);
 824
 825                 /*
 826                  * We can skip WAL-logging the insertions, unless PITR is in use.
 827                  *
 828                  * Note that for a non-temp INTO table, this is safe only because we
 829                  * know that the catalog changes above will have been WAL-logged, and
 830                  * so RecordTransactionCommit will think it needs to WAL-log the
 831                  * eventual transaction commit.  Else the commit might be lost, even
 832                  * though all the data is safely fsync'd ...
 833                  */
 834                 estate->es_into_relation_use_wal = XLogArchivingActive();
 835         }
 836
 837         estate->es_into_relation_descriptor = intoRelationDesc;
 838
 839         queryDesc->tupDesc = tupType;
 840         queryDesc->planstate = planstate;
 841 }
 842
 843 /*
 844  * Initialize ResultRelInfo data for one result relation
 845  */
 846 static void
 847 initResultRelInfo(ResultRelInfo *resultRelInfo,
 848                                   Index resultRelationIndex,
 849                                   List *rangeTable,
 850                                   CmdType operation,
 851                                   bool doInstrument)
 852 {
 853         Oid                     resultRelationOid;
 854         Relation        resultRelationDesc;
 855
 856         resultRelationOid = getrelid(resultRelationIndex, rangeTable);
 857         resultRelationDesc = heap_open(resultRelationOid, RowExclusiveLock);
 858
 859         switch (resultRelationDesc->rd_rel->relkind)
 860         {
 861                 case RELKIND_SEQUENCE:
 862                         ereport(ERROR,
 863                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 864                                          errmsg("cannot change sequence \"%s\"",
 865                                                         RelationGetRelationName(resultRelationDesc))));
 866                         break;
 867                 case RELKIND_TOASTVALUE:
 868                         ereport(ERROR,
 869                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 870                                          errmsg("cannot change TOAST relation \"%s\"",
 871                                                         RelationGetRelationName(resultRelationDesc))));
 872                         break;
 873                 case RELKIND_VIEW:
 874                         ereport(ERROR,
 875                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 876                                          errmsg("cannot change view \"%s\"",
 877                                                         RelationGetRelationName(resultRelationDesc))));
 878                         break;
 879         }
 880
 881         MemSet(resultRelInfo, 0, sizeof(ResultRelInfo));
 882         resultRelInfo->type = T_ResultRelInfo;
 883         resultRelInfo->ri_RangeTableIndex = resultRelationIndex;
 884         resultRelInfo->ri_RelationDesc = resultRelationDesc;
 885         resultRelInfo->ri_NumIndices = 0;
 886         resultRelInfo->ri_IndexRelationDescs = NULL;
 887         resultRelInfo->ri_IndexRelationInfo = NULL;
 888         /* make a copy so as not to depend on relcache info not changing... */
 889         resultRelInfo->ri_TrigDesc = CopyTriggerDesc(resultRelationDesc->trigdesc);
 890         if (resultRelInfo->ri_TrigDesc)
 891         {
 892                 int                     n = resultRelInfo->ri_TrigDesc->numtriggers;
 893
 894                 resultRelInfo->ri_TrigFunctions = (FmgrInfo *)
 895                         palloc0(n * sizeof(FmgrInfo));
 896                 if (doInstrument)
 897                         resultRelInfo->ri_TrigInstrument = InstrAlloc(n);
 898                 else
 899                         resultRelInfo->ri_TrigInstrument = NULL;
 900         }
 901         else
 902         {
 903                 resultRelInfo->ri_TrigFunctions = NULL;
 904                 resultRelInfo->ri_TrigInstrument = NULL;
 905         }
 906         resultRelInfo->ri_ConstraintExprs = NULL;
 907         resultRelInfo->ri_junkFilter = NULL;
 908
 909         /*
 910          * If there are indices on the result relation, open them and save
 911          * descriptors in the result relation info, so that we can add new index
 912          * entries for the tuples we add/update.  We need not do this for a
 913          * DELETE, however, since deletion doesn't affect indexes.
 914          */
 915         if (resultRelationDesc->rd_rel->relhasindex &&
 916                 operation != CMD_DELETE)
 917                 ExecOpenIndices(resultRelInfo);
 918 }
 919
 920 /*
 921  *              ExecContextForcesOids
 922  *
 923  * This is pretty grotty: when doing INSERT, UPDATE, or SELECT INTO,
 924  * we need to ensure that result tuples have space for an OID iff they are
 925  * going to be stored into a relation that has OIDs.  In other contexts
 926  * we are free to choose whether to leave space for OIDs in result tuples
 927  * (we generally don't want to, but we do if a physical-tlist optimization
 928  * is possible).  This routine checks the plan context and returns TRUE if the
 929  * choice is forced, FALSE if the choice is not forced.  In the TRUE case,
 930  * *hasoids is set to the required value.
 931  *
 932  * One reason this is ugly is that all plan nodes in the plan tree will emit
 933  * tuples with space for an OID, though we really only need the topmost node
 934  * to do so.  However, node types like Sort don't project new tuples but just
 935  * return their inputs, and in those cases the requirement propagates down
 936  * to the input node.  Eventually we might make this code smart enough to
 937  * recognize how far down the requirement really goes, but for now we just
 938  * make all plan nodes do the same thing if the top level forces the choice.
 939  *
 940  * We assume that estate->es_result_relation_info is already set up to
 941  * describe the target relation.  Note that in an UPDATE that spans an
 942  * inheritance tree, some of the target relations may have OIDs and some not.
 943  * We have to make the decisions on a per-relation basis as we initialize
 944  * each of the child plans of the topmost Append plan.
 945  *
 946  * SELECT INTO is even uglier, because we don't have the INTO relation's
 947  * descriptor available when this code runs; we have to look aside at a
 948  * flag set by InitPlan().
 949  */
 950 bool
 951 ExecContextForcesOids(PlanState *planstate, bool *hasoids)
 952 {
 953         if (planstate->state->es_select_into)
 954         {
 955                 *hasoids = planstate->state->es_into_oids;
 956                 return true;
 957         }
 958         else
 959         {
 960                 ResultRelInfo *ri = planstate->state->es_result_relation_info;
 961
 962                 if (ri != NULL)
 963                 {
 964                         Relation        rel = ri->ri_RelationDesc;
 965
 966                         if (rel != NULL)
 967                         {
 968                                 *hasoids = rel->rd_rel->relhasoids;
 969                                 return true;
 970                         }
 971                 }
 972         }
 973
 974         return false;
 975 }
 976
 977 /* ----------------------------------------------------------------
 978  *              ExecEndPlan
 979  *
 980  *              Cleans up the query plan -- closes files and frees up storage
 981  *
 982  * NOTE: we are no longer very worried about freeing storage per se
 983  * in this code; FreeExecutorState should be guaranteed to release all
 984  * memory that needs to be released.  What we are worried about doing
 985  * is closing relations and dropping buffer pins.  Thus, for example,
 986  * tuple tables must be cleared or dropped to ensure pins are released.
 987  * ----------------------------------------------------------------
 988  */
 989 void
 990 ExecEndPlan(PlanState *planstate, EState *estate)
 991 {
 992         ResultRelInfo *resultRelInfo;
 993         int                     i;
 994         ListCell   *l;
 995
 996         /*
 997          * shut down any PlanQual processing we were doing
 998          */
 999         if (estate->es_evalPlanQual != NULL)
1000                 EndEvalPlanQual(estate);
1001
1002         /*
1003          * shut down the node-type-specific query processing
1004          */
1005         ExecEndNode(planstate);
1006
1007         /*
1008          * destroy the executor "tuple" table.
1009          */
1010         ExecDropTupleTable(estate->es_tupleTable, true);
1011         estate->es_tupleTable = NULL;
1012
1013         /*
1014          * close the result relation(s) if any, but hold locks until xact commit.
1015          */
1016         resultRelInfo = estate->es_result_relations;
1017         for (i = estate->es_num_result_relations; i > 0; i--)
1018         {
1019                 /* Close indices and then the relation itself */
1020                 ExecCloseIndices(resultRelInfo);
1021                 heap_close(resultRelInfo->ri_RelationDesc, NoLock);
1022                 resultRelInfo++;
1023         }
1024
1025         /*
1026          * close the "into" relation if necessary, again keeping lock
1027          */
1028         if (estate->es_into_relation_descriptor != NULL)
1029         {
1030                 /*
1031                  * If we skipped using WAL, and it's not a temp relation, we must
1032                  * force the relation down to disk before it's safe to commit the
1033                  * transaction.  This requires forcing out any dirty buffers and then
1034                  * doing a forced fsync.
1035                  */
1036                 if (!estate->es_into_relation_use_wal &&
1037                         !estate->es_into_relation_descriptor->rd_istemp)
1038                 {
1039                         FlushRelationBuffers(estate->es_into_relation_descriptor);
1040                         /* FlushRelationBuffers will have opened rd_smgr */
1041                         smgrimmedsync(estate->es_into_relation_descriptor->rd_smgr);
1042                 }
1043
1044                 heap_close(estate->es_into_relation_descriptor, NoLock);
1045         }
1046
1047         /*
1048          * close any relations selected FOR UPDATE/FOR SHARE, again keeping locks
1049          */
1050         foreach(l, estate->es_rowMarks)
1051         {
1052                 ExecRowMark *erm = lfirst(l);
1053
1054                 heap_close(erm->relation, NoLock);
1055         }
1056 }
1057
1058 /* ----------------------------------------------------------------
1059  *              ExecutePlan
1060  *
1061  *              processes the query plan to retrieve 'numberTuples' tuples in the
1062  *              direction specified.
1063  *
1064  *              Retrieves all tuples if numberTuples is 0
1065  *
1066  *              result is either a slot containing the last tuple in the case
1067  *              of a SELECT or NULL otherwise.
1068  *
1069  * Note: the ctid attribute is a 'junk' attribute that is removed before the
1070  * user can see it
1071  * ----------------------------------------------------------------
1072  */
1073 static TupleTableSlot *
1074 ExecutePlan(EState *estate,
1075                         PlanState *planstate,
1076                         CmdType operation,
1077                         long numberTuples,
1078                         ScanDirection direction,
1079                         DestReceiver *dest)
1080 {
1081         JunkFilter *junkfilter;
1082         TupleTableSlot *slot;
1083         ItemPointer tupleid = NULL;
1084         ItemPointerData tuple_ctid;
1085         long            current_tuple_count;
1086         TupleTableSlot *result;
1087
1088         /*
1089          * initialize local variables
1090          */
1091         slot = NULL;
1092         current_tuple_count = 0;
1093         result = NULL;
1094
1095         /*
1096          * Set the direction.
1097          */
1098         estate->es_direction = direction;
1099
1100         /*
1101          * Process BEFORE EACH STATEMENT triggers
1102          */
1103         switch (operation)
1104         {
1105                 case CMD_UPDATE:
1106                         ExecBSUpdateTriggers(estate, estate->es_result_relation_info);
1107                         break;
1108                 case CMD_DELETE:
1109                         ExecBSDeleteTriggers(estate, estate->es_result_relation_info);
1110                         break;
1111                 case CMD_INSERT:
1112                         ExecBSInsertTriggers(estate, estate->es_result_relation_info);
1113                         break;
1114                 default:
1115                         /* do nothing */
1116                         break;
1117         }
1118
1119         /*
1120          * Loop until we've processed the proper number of tuples from the plan.
1121          */
1122
1123         for (;;)
1124         {
1125                 /* Reset the per-output-tuple exprcontext */
1126                 ResetPerTupleExprContext(estate);
1127
1128                 /*
1129                  * Execute the plan and obtain a tuple
1130                  */
1131 lnext:  ;
1132                 if (estate->es_useEvalPlan)
1133                 {
1134                         slot = EvalPlanQualNext(estate);
1135                         if (TupIsNull(slot))
1136                                 slot = ExecProcNode(planstate);
1137                 }
1138                 else
1139                         slot = ExecProcNode(planstate);
1140
1141                 /*
1142                  * if the tuple is null, then we assume there is nothing more to
1143                  * process so we just return null...
1144                  */
1145                 if (TupIsNull(slot))
1146                 {
1147                         result = NULL;
1148                         break;
1149                 }
1150
1151                 /*
1152                  * if we have a junk filter, then project a new tuple with the junk
1153                  * removed.
1154                  *
1155                  * Store this new "clean" tuple in the junkfilter's resultSlot.
1156                  * (Formerly, we stored it back over the "dirty" tuple, which is WRONG
1157                  * because that tuple slot has the wrong descriptor.)
1158                  *
1159                  * Also, extract all the junk information we need.
1160                  */
1161                 if ((junkfilter = estate->es_junkFilter) != NULL)
1162                 {
1163                         Datum           datum;
1164                         bool            isNull;
1165
1166                         /*
1167                          * extract the 'ctid' junk attribute.
1168                          */
1169                         if (operation == CMD_UPDATE || operation == CMD_DELETE)
1170                         {
1171                                 if (!ExecGetJunkAttribute(junkfilter,
1172                                                                                   slot,
1173                                                                                   "ctid",
1174                                                                                   &datum,
1175                                                                                   &isNull))
1176                                         elog(ERROR, "could not find junk ctid column");
1177
1178                                 /* shouldn't ever get a null result... */
1179                                 if (isNull)
1180                                         elog(ERROR, "ctid is NULL");
1181
1182                                 tupleid = (ItemPointer) DatumGetPointer(datum);
1183                                 tuple_ctid = *tupleid;  /* make sure we don't free the ctid!! */
1184                                 tupleid = &tuple_ctid;
1185                         }
1186
1187                         /*
1188                          * Process any FOR UPDATE or FOR SHARE locking requested.
1189                          */
1190                         else if (estate->es_rowMarks != NIL)
1191                         {
1192                                 ListCell   *l;
1193
1194                 lmark:  ;
1195                                 foreach(l, estate->es_rowMarks)
1196                                 {
1197                                         ExecRowMark *erm = lfirst(l);
1198                                         HeapTupleData tuple;
1199                                         Buffer          buffer;
1200                                         ItemPointerData update_ctid;
1201                                         TransactionId update_xmax;
1202                                         TupleTableSlot *newSlot;
1203                                         LockTupleMode lockmode;
1204                                         HTSU_Result test;
1205
1206                                         if (!ExecGetJunkAttribute(junkfilter,
1207                                                                                           slot,
1208                                                                                           erm->resname,
1209                                                                                           &datum,
1210                                                                                           &isNull))
1211                                                 elog(ERROR, "could not find junk \"%s\" column",
1212                                                          erm->resname);
1213
1214                                         /* shouldn't ever get a null result... */
1215                                         if (isNull)
1216                                                 elog(ERROR, "\"%s\" is NULL", erm->resname);
1217
1218                                         tuple.t_self = *((ItemPointer) DatumGetPointer(datum));
1219
1220                                         if (erm->forUpdate)
1221                                                 lockmode = LockTupleExclusive;
1222                                         else
1223                                                 lockmode = LockTupleShared;
1224
1225                                         test = heap_lock_tuple(erm->relation, &tuple, &buffer,
1226                                                                                    &update_ctid, &update_xmax,
1227                                                                                    estate->es_snapshot->curcid,
1228                                                                                    lockmode, erm->noWait);
1229                                         ReleaseBuffer(buffer);
1230                                         switch (test)
1231                                         {
1232                                                 case HeapTupleSelfUpdated:
1233                                                         /* treat it as deleted; do not process */
1234                                                         goto lnext;
1235
1236                                                 case HeapTupleMayBeUpdated:
1237                                                         break;
1238
1239                                                 case HeapTupleUpdated:
1240                                                         if (IsXactIsoLevelSerializable)
1241                                                                 ereport(ERROR,
1242                                                                  (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1243                                                                   errmsg("could not serialize access due to concurrent update")));
1244                                                         if (!ItemPointerEquals(&update_ctid,
1245                                                                                                    &tuple.t_self))
1246                                                         {
1247                                                                 /* updated, so look at updated version */
1248                                                                 newSlot = EvalPlanQual(estate,
1249                                                                                                            erm->rti,
1250                                                                                                            &update_ctid,
1251                                                                                                            update_xmax,
1252                                                                                                            estate->es_snapshot->curcid);
1253                                                                 if (!TupIsNull(newSlot))
1254                                                                 {
1255                                                                         slot = newSlot;
1256                                                                         estate->es_useEvalPlan = true;
1257                                                                         goto lmark;
1258                                                                 }
1259                                                         }
1260
1261                                                         /*
1262                                                          * if tuple was deleted or PlanQual failed for
1263                                                          * updated tuple - we must not return this tuple!
1264                                                          */
1265                                                         goto lnext;
1266
1267                                                 default:
1268                                                         elog(ERROR, "unrecognized heap_lock_tuple status: %u",
1269                                                                  test);
1270                                                         return NULL;
1271                                         }
1272                                 }
1273                         }
1274
1275                         /*
1276                          * Finally create a new "clean" tuple with all junk attributes
1277                          * removed
1278                          */
1279                         slot = ExecFilterJunk(junkfilter, slot);
1280                 }
1281
1282                 /*
1283                  * now that we have a tuple, do the appropriate thing with it.. either
1284                  * return it to the user, add it to a relation someplace, delete it
1285                  * from a relation, or modify some of its attributes.
1286                  */
1287                 switch (operation)
1288                 {
1289                         case CMD_SELECT:
1290                                 ExecSelect(slot,        /* slot containing tuple */
1291                                                    dest,        /* destination's tuple-receiver obj */
1292                                                    estate);
1293                                 result = slot;
1294                                 break;
1295
1296                         case CMD_INSERT:
1297                                 ExecInsert(slot, tupleid, estate);
1298                                 result = NULL;
1299                                 break;
1300
1301                         case CMD_DELETE:
1302                                 ExecDelete(slot, tupleid, estate);
1303                                 result = NULL;
1304                                 break;
1305
1306                         case CMD_UPDATE:
1307                                 ExecUpdate(slot, tupleid, estate);
1308                                 result = NULL;
1309                                 break;
1310
1311                         default:
1312                                 elog(ERROR, "unrecognized operation code: %d",
1313                                          (int) operation);
1314                                 result = NULL;
1315                                 break;
1316                 }
1317
1318                 /*
1319                  * check our tuple count.. if we've processed the proper number then
1320                  * quit, else loop again and process more tuples.  Zero numberTuples
1321                  * means no limit.
1322                  */
1323                 current_tuple_count++;
1324                 if (numberTuples && numberTuples == current_tuple_count)
1325                         break;
1326         }
1327
1328         /*
1329          * Process AFTER EACH STATEMENT triggers
1330          */
1331         switch (operation)
1332         {
1333                 case CMD_UPDATE:
1334                         ExecASUpdateTriggers(estate, estate->es_result_relation_info);
1335                         break;
1336                 case CMD_DELETE:
1337                         ExecASDeleteTriggers(estate, estate->es_result_relation_info);
1338                         break;
1339                 case CMD_INSERT:
1340                         ExecASInsertTriggers(estate, estate->es_result_relation_info);
1341                         break;
1342                 default:
1343                         /* do nothing */
1344                         break;
1345         }
1346
1347         /*
1348          * here, result is either a slot containing a tuple in the case of a
1349          * SELECT or NULL otherwise.
1350          */
1351         return result;
1352 }
1353
1354 /* ----------------------------------------------------------------
1355  *              ExecSelect
1356  *
1357  *              SELECTs are easy.. we just pass the tuple to the appropriate
1358  *              print function.  The only complexity is when we do a
1359  *              "SELECT INTO", in which case we insert the tuple into
1360  *              the appropriate relation (note: this is a newly created relation
1361  *              so we don't need to worry about indices or locks.)
1362  * ----------------------------------------------------------------
1363  */
1364 static void
1365 ExecSelect(TupleTableSlot *slot,
1366                    DestReceiver *dest,
1367                    EState *estate)
1368 {
1369         /*
1370          * insert the tuple into the "into relation"
1371          *
1372          * XXX this probably ought to be replaced by a separate destination
1373          */
1374         if (estate->es_into_relation_descriptor != NULL)
1375         {
1376                 HeapTuple       tuple;
1377
1378                 tuple = ExecCopySlotTuple(slot);
1379                 heap_insert(estate->es_into_relation_descriptor, tuple,
1380                                         estate->es_snapshot->curcid,
1381                                         estate->es_into_relation_use_wal,
1382                                         false);         /* never any point in using FSM */
1383                 /* we know there are no indexes to update */
1384                 heap_freetuple(tuple);
1385                 IncrAppended();
1386         }
1387
1388         /*
1389          * send the tuple to the destination
1390          */
1391         (*dest->receiveSlot) (slot, dest);
1392         IncrRetrieved();
1393         (estate->es_processed)++;
1394 }
1395
1396 /* ----------------------------------------------------------------
1397  *              ExecInsert
1398  *
1399  *              INSERTs are trickier.. we have to insert the tuple into
1400  *              the base relation and insert appropriate tuples into the
1401  *              index relations.
1402  * ----------------------------------------------------------------
1403  */
1404 static void
1405 ExecInsert(TupleTableSlot *slot,
1406                    ItemPointer tupleid,
1407                    EState *estate)
1408 {
1409         HeapTuple       tuple;
1410         ResultRelInfo *resultRelInfo;
1411         Relation        resultRelationDesc;
1412         Oid                     newId;
1413
1414         /*
1415          * get the heap tuple out of the tuple table slot, making sure we have a
1416          * writable copy
1417          */
1418         tuple = ExecMaterializeSlot(slot);
1419
1420         /*
1421          * get information on the (current) result relation
1422          */
1423         resultRelInfo = estate->es_result_relation_info;
1424         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1425
1426         /* BEFORE ROW INSERT Triggers */
1427         if (resultRelInfo->ri_TrigDesc &&
1428                 resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_INSERT] > 0)
1429         {
1430                 HeapTuple       newtuple;
1431
1432                 newtuple = ExecBRInsertTriggers(estate, resultRelInfo, tuple);
1433
1434                 if (newtuple == NULL)   /* "do nothing" */
1435                         return;
1436
1437                 if (newtuple != tuple)  /* modified by Trigger(s) */
1438                 {
1439                         /*
1440                          * Put the modified tuple into a slot for convenience of routines
1441                          * below.  We assume the tuple was allocated in per-tuple memory
1442                          * context, and therefore will go away by itself. The tuple table
1443                          * slot should not try to clear it.
1444                          */
1445                         TupleTableSlot *newslot = estate->es_trig_tuple_slot;
1446
1447                         if (newslot->tts_tupleDescriptor != slot->tts_tupleDescriptor)
1448                                 ExecSetSlotDescriptor(newslot,
1449                                                                           slot->tts_tupleDescriptor,
1450                                                                           false);
1451                         ExecStoreTuple(newtuple, newslot, InvalidBuffer, false);
1452                         slot = newslot;
1453                         tuple = newtuple;
1454                 }
1455         }
1456
1457         /*
1458          * Check the constraints of the tuple
1459          */
1460         if (resultRelationDesc->rd_att->constr)
1461                 ExecConstraints(resultRelInfo, slot, estate);
1462
1463         /*
1464          * insert the tuple
1465          *
1466          * Note: heap_insert returns the tid (location) of the new tuple in the
1467          * t_self field.
1468          */
1469         newId = heap_insert(resultRelationDesc, tuple,
1470                                                 estate->es_snapshot->curcid,
1471                                                 true, true);
1472
1473         IncrAppended();
1474         (estate->es_processed)++;
1475         estate->es_lastoid = newId;
1476         setLastTid(&(tuple->t_self));
1477
1478         /*
1479          * insert index entries for tuple
1480          */
1481         if (resultRelInfo->ri_NumIndices > 0)
1482                 ExecInsertIndexTuples(slot, &(tuple->t_self), estate, false);
1483
1484         /* AFTER ROW INSERT Triggers */
1485         ExecARInsertTriggers(estate, resultRelInfo, tuple);
1486 }
1487
1488 /* ----------------------------------------------------------------
1489  *              ExecDelete
1490  *
1491  *              DELETE is like UPDATE, except that we delete the tuple and no
1492  *              index modifications are needed
1493  * ----------------------------------------------------------------
1494  */
1495 static void
1496 ExecDelete(TupleTableSlot *slot,
1497                    ItemPointer tupleid,
1498                    EState *estate)
1499 {
1500         ResultRelInfo *resultRelInfo;
1501         Relation        resultRelationDesc;
1502         HTSU_Result result;
1503         ItemPointerData update_ctid;
1504         TransactionId update_xmax;
1505
1506         /*
1507          * get information on the (current) result relation
1508          */
1509         resultRelInfo = estate->es_result_relation_info;
1510         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1511
1512         /* BEFORE ROW DELETE Triggers */
1513         if (resultRelInfo->ri_TrigDesc &&
1514                 resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_DELETE] > 0)
1515         {
1516                 bool            dodelete;
1517
1518                 dodelete = ExecBRDeleteTriggers(estate, resultRelInfo, tupleid,
1519                                                                                 estate->es_snapshot->curcid);
1520
1521                 if (!dodelete)                  /* "do nothing" */
1522                         return;
1523         }
1524
1525         /*
1526          * delete the tuple
1527          *
1528          * Note: if es_crosscheck_snapshot isn't InvalidSnapshot, we check that
1529          * the row to be deleted is visible to that snapshot, and throw a can't-
1530          * serialize error if not.      This is a special-case behavior needed for
1531          * referential integrity updates in serializable transactions.
1532          */
1533 ldelete:;
1534         result = heap_delete(resultRelationDesc, tupleid,
1535                                                  &update_ctid, &update_xmax,
1536                                                  estate->es_snapshot->curcid,
1537                                                  estate->es_crosscheck_snapshot,
1538                                                  true /* wait for commit */ );
1539         switch (result)
1540         {
1541                 case HeapTupleSelfUpdated:
1542                         /* already deleted by self; nothing to do */
1543                         return;
1544
1545                 case HeapTupleMayBeUpdated:
1546                         break;
1547
1548                 case HeapTupleUpdated:
1549                         if (IsXactIsoLevelSerializable)
1550                                 ereport(ERROR,
1551                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1552                                                  errmsg("could not serialize access due to concurrent update")));
1553                         else if (!ItemPointerEquals(tupleid, &update_ctid))
1554                         {
1555                                 TupleTableSlot *epqslot;
1556
1557                                 epqslot = EvalPlanQual(estate,
1558                                                                            resultRelInfo->ri_RangeTableIndex,
1559                                                                            &update_ctid,
1560                                                                            update_xmax,
1561                                                                            estate->es_snapshot->curcid);
1562                                 if (!TupIsNull(epqslot))
1563                                 {
1564                                         *tupleid = update_ctid;
1565                                         goto ldelete;
1566                                 }
1567                         }
1568                         /* tuple already deleted; nothing to do */
1569                         return;
1570
1571                 default:
1572                         elog(ERROR, "unrecognized heap_delete status: %u", result);
1573                         return;
1574         }
1575
1576         IncrDeleted();
1577         (estate->es_processed)++;
1578
1579         /*
1580          * Note: Normally one would think that we have to delete index tuples
1581          * associated with the heap tuple now...
1582          *
1583          * ... but in POSTGRES, we have no need to do this because VACUUM will
1584          * take care of it later.  We can't delete index tuples immediately
1585          * anyway, since the tuple is still visible to other transactions.
1586          */
1587
1588         /* AFTER ROW DELETE Triggers */
1589         ExecARDeleteTriggers(estate, resultRelInfo, tupleid);
1590 }
1591
1592 /* ----------------------------------------------------------------
1593  *              ExecUpdate
1594  *
1595  *              note: we can't run UPDATE queries with transactions
1596  *              off because UPDATEs are actually INSERTs and our
1597  *              scan will mistakenly loop forever, updating the tuple
1598  *              it just inserted..      This should be fixed but until it
1599  *              is, we don't want to get stuck in an infinite loop
1600  *              which corrupts your database..
1601  * ----------------------------------------------------------------
1602  */
1603 static void
1604 ExecUpdate(TupleTableSlot *slot,
1605                    ItemPointer tupleid,
1606                    EState *estate)
1607 {
1608         HeapTuple       tuple;
1609         ResultRelInfo *resultRelInfo;
1610         Relation        resultRelationDesc;
1611         HTSU_Result result;
1612         ItemPointerData update_ctid;
1613         TransactionId update_xmax;
1614
1615         /*
1616          * abort the operation if not running transactions
1617          */
1618         if (IsBootstrapProcessingMode())
1619                 elog(ERROR, "cannot UPDATE during bootstrap");
1620
1621         /*
1622          * get the heap tuple out of the tuple table slot, making sure we have a
1623          * writable copy
1624          */
1625         tuple = ExecMaterializeSlot(slot);
1626
1627         /*
1628          * get information on the (current) result relation
1629          */
1630         resultRelInfo = estate->es_result_relation_info;
1631         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1632
1633         /* BEFORE ROW UPDATE Triggers */
1634         if (resultRelInfo->ri_TrigDesc &&
1635                 resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_UPDATE] > 0)
1636         {
1637                 HeapTuple       newtuple;
1638
1639                 newtuple = ExecBRUpdateTriggers(estate, resultRelInfo,
1640                                                                                 tupleid, tuple,
1641                                                                                 estate->es_snapshot->curcid);
1642
1643                 if (newtuple == NULL)   /* "do nothing" */
1644                         return;
1645
1646                 if (newtuple != tuple)  /* modified by Trigger(s) */
1647                 {
1648                         /*
1649                          * Put the modified tuple into a slot for convenience of routines
1650                          * below.  We assume the tuple was allocated in per-tuple memory
1651                          * context, and therefore will go away by itself. The tuple table
1652                          * slot should not try to clear it.
1653                          */
1654                         TupleTableSlot *newslot = estate->es_trig_tuple_slot;
1655
1656                         if (newslot->tts_tupleDescriptor != slot->tts_tupleDescriptor)
1657                                 ExecSetSlotDescriptor(newslot,
1658                                                                           slot->tts_tupleDescriptor,
1659                                                                           false);
1660                         ExecStoreTuple(newtuple, newslot, InvalidBuffer, false);
1661                         slot = newslot;
1662                         tuple = newtuple;
1663                 }
1664         }
1665
1666         /*
1667          * Check the constraints of the tuple
1668          *
1669          * If we generate a new candidate tuple after EvalPlanQual testing, we
1670          * must loop back here and recheck constraints.  (We don't need to redo
1671          * triggers, however.  If there are any BEFORE triggers then trigger.c
1672          * will have done heap_lock_tuple to lock the correct tuple, so there's no
1673          * need to do them again.)
1674          */
1675 lreplace:;
1676         if (resultRelationDesc->rd_att->constr)
1677                 ExecConstraints(resultRelInfo, slot, estate);
1678
1679         /*
1680          * replace the heap tuple
1681          *
1682          * Note: if es_crosscheck_snapshot isn't InvalidSnapshot, we check that
1683          * the row to be updated is visible to that snapshot, and throw a can't-
1684          * serialize error if not.      This is a special-case behavior needed for
1685          * referential integrity updates in serializable transactions.
1686          */
1687         result = heap_update(resultRelationDesc, tupleid, tuple,
1688                                                  &update_ctid, &update_xmax,
1689                                                  estate->es_snapshot->curcid,
1690                                                  estate->es_crosscheck_snapshot,
1691                                                  true /* wait for commit */ );
1692         switch (result)
1693         {
1694                 case HeapTupleSelfUpdated:
1695                         /* already deleted by self; nothing to do */
1696                         return;
1697
1698                 case HeapTupleMayBeUpdated:
1699                         break;
1700
1701                 case HeapTupleUpdated:
1702                         if (IsXactIsoLevelSerializable)
1703                                 ereport(ERROR,
1704                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1705                                                  errmsg("could not serialize access due to concurrent update")));
1706                         else if (!ItemPointerEquals(tupleid, &update_ctid))
1707                         {
1708                                 TupleTableSlot *epqslot;
1709
1710                                 epqslot = EvalPlanQual(estate,
1711                                                                            resultRelInfo->ri_RangeTableIndex,
1712                                                                            &update_ctid,
1713                                                                            update_xmax,
1714                                                                            estate->es_snapshot->curcid);
1715                                 if (!TupIsNull(epqslot))
1716                                 {
1717                                         *tupleid = update_ctid;
1718                                         slot = ExecFilterJunk(estate->es_junkFilter, epqslot);
1719                                         tuple = ExecMaterializeSlot(slot);
1720                                         goto lreplace;
1721                                 }
1722                         }
1723                         /* tuple already deleted; nothing to do */
1724                         return;
1725
1726                 default:
1727                         elog(ERROR, "unrecognized heap_update status: %u", result);
1728                         return;
1729         }
1730
1731         IncrReplaced();
1732         (estate->es_processed)++;
1733
1734         /*
1735          * Note: instead of having to update the old index tuples associated with
1736          * the heap tuple, all we do is form and insert new index tuples. This is
1737          * because UPDATEs are actually DELETEs and INSERTs, and index tuple
1738          * deletion is done later by VACUUM (see notes in ExecDelete).  All we do
1739          * here is insert new index tuples.  -cim 9/27/89
1740          */
1741
1742         /*
1743          * insert index entries for tuple
1744          *
1745          * Note: heap_update returns the tid (location) of the new tuple in the
1746          * t_self field.
1747          */
1748         if (resultRelInfo->ri_NumIndices > 0)
1749                 ExecInsertIndexTuples(slot, &(tuple->t_self), estate, false);
1750
1751         /* AFTER ROW UPDATE Triggers */
1752         ExecARUpdateTriggers(estate, resultRelInfo, tupleid, tuple);
1753 }
1754
1755 static const char *
1756 ExecRelCheck(ResultRelInfo *resultRelInfo,
1757                          TupleTableSlot *slot, EState *estate)
1758 {
1759         Relation        rel = resultRelInfo->ri_RelationDesc;
1760         int                     ncheck = rel->rd_att->constr->num_check;
1761         ConstrCheck *check = rel->rd_att->constr->check;
1762         ExprContext *econtext;
1763         MemoryContext oldContext;
1764         List       *qual;
1765         int                     i;
1766
1767         /*
1768          * If first time through for this result relation, build expression
1769          * nodetrees for rel's constraint expressions.  Keep them in the per-query
1770          * memory context so they'll survive throughout the query.
1771          */
1772         if (resultRelInfo->ri_ConstraintExprs == NULL)
1773         {
1774                 oldContext = MemoryContextSwitchTo(estate->es_query_cxt);
1775                 resultRelInfo->ri_ConstraintExprs =
1776                         (List **) palloc(ncheck * sizeof(List *));
1777                 for (i = 0; i < ncheck; i++)
1778                 {
1779                         /* ExecQual wants implicit-AND form */
1780                         qual = make_ands_implicit(stringToNode(check[i].ccbin));
1781                         resultRelInfo->ri_ConstraintExprs[i] = (List *)
1782                                 ExecPrepareExpr((Expr *) qual, estate);
1783                 }
1784                 MemoryContextSwitchTo(oldContext);
1785         }
1786
1787         /*
1788          * We will use the EState's per-tuple context for evaluating constraint
1789          * expressions (creating it if it's not already there).
1790          */
1791         econtext = GetPerTupleExprContext(estate);
1792
1793         /* Arrange for econtext's scan tuple to be the tuple under test */
1794         econtext->ecxt_scantuple = slot;
1795
1796         /* And evaluate the constraints */
1797         for (i = 0; i < ncheck; i++)
1798         {
1799                 qual = resultRelInfo->ri_ConstraintExprs[i];
1800
1801                 /*
1802                  * NOTE: SQL92 specifies that a NULL result from a constraint
1803                  * expression is not to be treated as a failure.  Therefore, tell
1804                  * ExecQual to return TRUE for NULL.
1805                  */
1806                 if (!ExecQual(qual, econtext, true))
1807                         return check[i].ccname;
1808         }
1809
1810         /* NULL result means no error */
1811         return NULL;
1812 }
1813
1814 void
1815 ExecConstraints(ResultRelInfo *resultRelInfo,
1816                                 TupleTableSlot *slot, EState *estate)
1817 {
1818         Relation        rel = resultRelInfo->ri_RelationDesc;
1819         TupleConstr *constr = rel->rd_att->constr;
1820
1821         Assert(constr);
1822
1823         if (constr->has_not_null)
1824         {
1825                 int                     natts = rel->rd_att->natts;
1826                 int                     attrChk;
1827
1828                 for (attrChk = 1; attrChk <= natts; attrChk++)
1829                 {
1830                         if (rel->rd_att->attrs[attrChk - 1]->attnotnull &&
1831                                 slot_attisnull(slot, attrChk))
1832                                 ereport(ERROR,
1833                                                 (errcode(ERRCODE_NOT_NULL_VIOLATION),
1834                                                  errmsg("null value in column \"%s\" violates not-null constraint",
1835                                                 NameStr(rel->rd_att->attrs[attrChk - 1]->attname))));
1836                 }
1837         }
1838
1839         if (constr->num_check > 0)
1840         {
1841                 const char *failed;
1842
1843                 if ((failed = ExecRelCheck(resultRelInfo, slot, estate)) != NULL)
1844                         ereport(ERROR,
1845                                         (errcode(ERRCODE_CHECK_VIOLATION),
1846                                          errmsg("new row for relation \"%s\" violates check constraint \"%s\"",
1847                                                         RelationGetRelationName(rel), failed)));
1848         }
1849 }
1850
1851 /*
1852  * Check a modified tuple to see if we want to process its updated version
1853  * under READ COMMITTED rules.
1854  *
1855  * See backend/executor/README for some info about how this works.
1856  *
1857  *      estate - executor state data
1858  *      rti - rangetable index of table containing tuple
1859  *      *tid - t_ctid from the outdated tuple (ie, next updated version)
1860  *      priorXmax - t_xmax from the outdated tuple
1861  *      curCid - command ID of current command of my transaction
1862  *
1863  * *tid is also an output parameter: it's modified to hold the TID of the
1864  * latest version of the tuple (note this may be changed even on failure)
1865  *
1866  * Returns a slot containing the new candidate update/delete tuple, or
1867  * NULL if we determine we shouldn't process the row.
1868  */
1869 TupleTableSlot *
1870 EvalPlanQual(EState *estate, Index rti,
1871                          ItemPointer tid, TransactionId priorXmax, CommandId curCid)
1872 {
1873         evalPlanQual *epq;
1874         EState     *epqstate;
1875         Relation        relation;
1876         HeapTupleData tuple;
1877         HeapTuple       copyTuple = NULL;
1878         bool            endNode;
1879
1880         Assert(rti != 0);
1881
1882         /*
1883          * find relation containing target tuple
1884          */
1885         if (estate->es_result_relation_info != NULL &&
1886                 estate->es_result_relation_info->ri_RangeTableIndex == rti)
1887                 relation = estate->es_result_relation_info->ri_RelationDesc;
1888         else
1889         {
1890                 ListCell   *l;
1891
1892                 relation = NULL;
1893                 foreach(l, estate->es_rowMarks)
1894                 {
1895                         if (((ExecRowMark *) lfirst(l))->rti == rti)
1896                         {
1897                                 relation = ((ExecRowMark *) lfirst(l))->relation;
1898                                 break;
1899                         }
1900                 }
1901                 if (relation == NULL)
1902                         elog(ERROR, "could not find RowMark for RT index %u", rti);
1903         }
1904
1905         /*
1906          * fetch tid tuple
1907          *
1908          * Loop here to deal with updated or busy tuples
1909          */
1910         tuple.t_self = *tid;
1911         for (;;)
1912         {
1913                 Buffer          buffer;
1914
1915                 if (heap_fetch(relation, SnapshotDirty, &tuple, &buffer, true, NULL))
1916                 {
1917                         /*
1918                          * If xmin isn't what we're expecting, the slot must have been
1919                          * recycled and reused for an unrelated tuple.  This implies that
1920                          * the latest version of the row was deleted, so we need do
1921                          * nothing.  (Should be safe to examine xmin without getting
1922                          * buffer's content lock, since xmin never changes in an existing
1923                          * tuple.)
1924                          */
1925                         if (!TransactionIdEquals(HeapTupleHeaderGetXmin(tuple.t_data),
1926                                                                          priorXmax))
1927                         {
1928                                 ReleaseBuffer(buffer);
1929                                 return NULL;
1930                         }
1931
1932                         /* otherwise xmin should not be dirty... */
1933                         if (TransactionIdIsValid(SnapshotDirty->xmin))
1934                                 elog(ERROR, "t_xmin is uncommitted in tuple to be updated");
1935
1936                         /*
1937                          * If tuple is being updated by other transaction then we have to
1938                          * wait for its commit/abort.
1939                          */
1940                         if (TransactionIdIsValid(SnapshotDirty->xmax))
1941                         {
1942                                 ReleaseBuffer(buffer);
1943                                 XactLockTableWait(SnapshotDirty->xmax);
1944                                 continue;               /* loop back to repeat heap_fetch */
1945                         }
1946
1947                         /*
1948                          * If tuple was inserted by our own transaction, we have to check
1949                          * cmin against curCid: cmin >= curCid means our command cannot
1950                          * see the tuple, so we should ignore it.  Without this we are
1951                          * open to the "Halloween problem" of indefinitely re-updating
1952                          * the same tuple.  (We need not check cmax because
1953                          * HeapTupleSatisfiesDirty will consider a tuple deleted by
1954                          * our transaction dead, regardless of cmax.)  We just checked
1955                          * that priorXmax == xmin, so we can test that variable instead
1956                          * of doing HeapTupleHeaderGetXmin again.
1957                          */
1958                         if (TransactionIdIsCurrentTransactionId(priorXmax) &&
1959                                 HeapTupleHeaderGetCmin(tuple.t_data) >= curCid)
1960                         {
1961                                 ReleaseBuffer(buffer);
1962                                 return NULL;
1963                         }
1964
1965                         /*
1966                          * We got tuple - now copy it for use by recheck query.
1967                          */
1968                         copyTuple = heap_copytuple(&tuple);
1969                         ReleaseBuffer(buffer);
1970                         break;
1971                 }
1972
1973                 /*
1974                  * If the referenced slot was actually empty, the latest version of
1975                  * the row must have been deleted, so we need do nothing.
1976                  */
1977                 if (tuple.t_data == NULL)
1978                 {
1979                         ReleaseBuffer(buffer);
1980                         return NULL;
1981                 }
1982
1983                 /*
1984                  * As above, if xmin isn't what we're expecting, do nothing.
1985                  */
1986                 if (!TransactionIdEquals(HeapTupleHeaderGetXmin(tuple.t_data),
1987                                                                  priorXmax))
1988                 {
1989                         ReleaseBuffer(buffer);
1990                         return NULL;
1991                 }
1992
1993                 /*
1994                  * If we get here, the tuple was found but failed SnapshotDirty.
1995                  * Assuming the xmin is either a committed xact or our own xact (as it
1996                  * certainly should be if we're trying to modify the tuple), this must
1997                  * mean that the row was updated or deleted by either a committed xact
1998                  * or our own xact.  If it was deleted, we can ignore it; if it was
1999                  * updated then chain up to the next version and repeat the whole
2000                  * test.
2001                  *
2002                  * As above, it should be safe to examine xmax and t_ctid without the
2003                  * buffer content lock, because they can't be changing.
2004                  */
2005                 if (ItemPointerEquals(&tuple.t_self, &tuple.t_data->t_ctid))
2006                 {
2007                         /* deleted, so forget about it */
2008                         ReleaseBuffer(buffer);
2009                         return NULL;
2010                 }
2011
2012                 /* updated, so look at the updated row */
2013                 tuple.t_self = tuple.t_data->t_ctid;
2014                 /* updated row should have xmin matching this xmax */
2015                 priorXmax = HeapTupleHeaderGetXmax(tuple.t_data);
2016                 ReleaseBuffer(buffer);
2017                 /* loop back to fetch next in chain */
2018         }
2019
2020         /*
2021          * For UPDATE/DELETE we have to return tid of actual row we're executing
2022          * PQ for.
2023          */
2024         *tid = tuple.t_self;
2025
2026         /*
2027          * Need to run a recheck subquery.      Find or create a PQ stack entry.
2028          */
2029         epq = estate->es_evalPlanQual;
2030         endNode = true;
2031
2032         if (epq != NULL && epq->rti == 0)
2033         {
2034                 /* Top PQ stack entry is idle, so re-use it */
2035                 Assert(!(estate->es_useEvalPlan) && epq->next == NULL);
2036                 epq->rti = rti;
2037                 endNode = false;
2038         }
2039
2040         /*
2041          * If this is request for another RTE - Ra, - then we have to check wasn't
2042          * PlanQual requested for Ra already and if so then Ra' row was updated
2043          * again and we have to re-start old execution for Ra and forget all what
2044          * we done after Ra was suspended. Cool? -:))
2045          */
2046         if (epq != NULL && epq->rti != rti &&
2047                 epq->estate->es_evTuple[rti - 1] != NULL)
2048         {
2049                 do
2050                 {
2051                         evalPlanQual *oldepq;
2052
2053                         /* stop execution */
2054                         EvalPlanQualStop(epq);
2055                         /* pop previous PlanQual from the stack */
2056                         oldepq = epq->next;
2057                         Assert(oldepq && oldepq->rti != 0);
2058                         /* push current PQ to freePQ stack */
2059                         oldepq->free = epq;
2060                         epq = oldepq;
2061                         estate->es_evalPlanQual = epq;
2062                 } while (epq->rti != rti);
2063         }
2064
2065         /*
2066          * If we are requested for another RTE then we have to suspend execution
2067          * of current PlanQual and start execution for new one.
2068          */
2069         if (epq == NULL || epq->rti != rti)
2070         {
2071                 /* try to reuse plan used previously */
2072                 evalPlanQual *newepq = (epq != NULL) ? epq->free : NULL;
2073
2074                 if (newepq == NULL)             /* first call or freePQ stack is empty */
2075                 {
2076                         newepq = (evalPlanQual *) palloc0(sizeof(evalPlanQual));
2077                         newepq->free = NULL;
2078                         newepq->estate = NULL;
2079                         newepq->planstate = NULL;
2080                 }
2081                 else
2082                 {
2083                         /* recycle previously used PlanQual */
2084                         Assert(newepq->estate == NULL);
2085                         epq->free = NULL;
2086                 }
2087                 /* push current PQ to the stack */
2088                 newepq->next = epq;
2089                 epq = newepq;
2090                 estate->es_evalPlanQual = epq;
2091                 epq->rti = rti;
2092                 endNode = false;
2093         }
2094
2095         Assert(epq->rti == rti);
2096
2097         /*
2098          * Ok - we're requested for the same RTE.  Unfortunately we still have to
2099          * end and restart execution of the plan, because ExecReScan wouldn't
2100          * ensure that upper plan nodes would reset themselves.  We could make
2101          * that work if insertion of the target tuple were integrated with the
2102          * Param mechanism somehow, so that the upper plan nodes know that their
2103          * children's outputs have changed.
2104          *
2105          * Note that the stack of free evalPlanQual nodes is quite useless at the
2106          * moment, since it only saves us from pallocing/releasing the
2107          * evalPlanQual nodes themselves.  But it will be useful once we implement
2108          * ReScan instead of end/restart for re-using PlanQual nodes.
2109          */
2110         if (endNode)
2111         {
2112                 /* stop execution */
2113                 EvalPlanQualStop(epq);
2114         }
2115
2116         /*
2117          * Initialize new recheck query.
2118          *
2119          * Note: if we were re-using PlanQual plans via ExecReScan, we'd need to
2120          * instead copy down changeable state from the top plan (including
2121          * es_result_relation_info, es_junkFilter) and reset locally changeable
2122          * state in the epq (including es_param_exec_vals, es_evTupleNull).
2123          */
2124         EvalPlanQualStart(epq, estate, epq->next);
2125
2126         /*
2127          * free old RTE' tuple, if any, and store target tuple where relation's
2128          * scan node will see it
2129          */
2130         epqstate = epq->estate;
2131         if (epqstate->es_evTuple[rti - 1] != NULL)
2132                 heap_freetuple(epqstate->es_evTuple[rti - 1]);
2133         epqstate->es_evTuple[rti - 1] = copyTuple;
2134
2135         return EvalPlanQualNext(estate);
2136 }
2137
2138 static TupleTableSlot *
2139 EvalPlanQualNext(EState *estate)
2140 {
2141         evalPlanQual *epq = estate->es_evalPlanQual;
2142         MemoryContext oldcontext;
2143         TupleTableSlot *slot;
2144
2145         Assert(epq->rti != 0);
2146
2147 lpqnext:;
2148         oldcontext = MemoryContextSwitchTo(epq->estate->es_query_cxt);
2149         slot = ExecProcNode(epq->planstate);
2150         MemoryContextSwitchTo(oldcontext);
2151
2152         /*
2153          * No more tuples for this PQ. Continue previous one.
2154          */
2155         if (TupIsNull(slot))
2156         {
2157                 evalPlanQual *oldepq;
2158
2159                 /* stop execution */
2160                 EvalPlanQualStop(epq);
2161                 /* pop old PQ from the stack */
2162                 oldepq = epq->next;
2163                 if (oldepq == NULL)
2164                 {
2165                         /* this is the first (oldest) PQ - mark as free */
2166                         epq->rti = 0;
2167                         estate->es_useEvalPlan = false;
2168                         /* and continue Query execution */
2169                         return NULL;
2170                 }
2171                 Assert(oldepq->rti != 0);
2172                 /* push current PQ to freePQ stack */
2173                 oldepq->free = epq;
2174                 epq = oldepq;
2175                 estate->es_evalPlanQual = epq;
2176                 goto lpqnext;
2177         }
2178
2179         return slot;
2180 }
2181
2182 static void
2183 EndEvalPlanQual(EState *estate)
2184 {
2185         evalPlanQual *epq = estate->es_evalPlanQual;
2186
2187         if (epq->rti == 0)                      /* plans already shutdowned */
2188         {
2189                 Assert(epq->next == NULL);
2190                 return;
2191         }
2192
2193         for (;;)
2194         {
2195                 evalPlanQual *oldepq;
2196
2197                 /* stop execution */
2198                 EvalPlanQualStop(epq);
2199                 /* pop old PQ from the stack */
2200                 oldepq = epq->next;
2201                 if (oldepq == NULL)
2202                 {
2203                         /* this is the first (oldest) PQ - mark as free */
2204                         epq->rti = 0;
2205                         estate->es_useEvalPlan = false;
2206                         break;
2207                 }
2208                 Assert(oldepq->rti != 0);
2209                 /* push current PQ to freePQ stack */
2210                 oldepq->free = epq;
2211                 epq = oldepq;
2212                 estate->es_evalPlanQual = epq;
2213         }
2214 }
2215
2216 /*
2217  * Start execution of one level of PlanQual.
2218  *
2219  * This is a cut-down version of ExecutorStart(): we copy some state from
2220  * the top-level estate rather than initializing it fresh.
2221  */
2222 static void
2223 EvalPlanQualStart(evalPlanQual *epq, EState *estate, evalPlanQual *priorepq)
2224 {
2225         EState     *epqstate;
2226         int                     rtsize;
2227         MemoryContext oldcontext;
2228
2229         rtsize = list_length(estate->es_range_table);
2230
2231         epq->estate = epqstate = CreateExecutorState();
2232
2233         oldcontext = MemoryContextSwitchTo(epqstate->es_query_cxt);
2234
2235         /*
2236          * The epqstates share the top query's copy of unchanging state such as
2237          * the snapshot, rangetable, result-rel info, and external Param info.
2238          * They need their own copies of local state, including a tuple table,
2239          * es_param_exec_vals, etc.
2240          */
2241         epqstate->es_direction = ForwardScanDirection;
2242         epqstate->es_snapshot = estate->es_snapshot;
2243         epqstate->es_crosscheck_snapshot = estate->es_crosscheck_snapshot;
2244         epqstate->es_range_table = estate->es_range_table;
2245         epqstate->es_result_relations = estate->es_result_relations;
2246         epqstate->es_num_result_relations = estate->es_num_result_relations;
2247         epqstate->es_result_relation_info = estate->es_result_relation_info;
2248         epqstate->es_junkFilter = estate->es_junkFilter;
2249         epqstate->es_into_relation_descriptor = estate->es_into_relation_descriptor;
2250         epqstate->es_into_relation_use_wal = estate->es_into_relation_use_wal;
2251         epqstate->es_param_list_info = estate->es_param_list_info;
2252         if (estate->es_topPlan->nParamExec > 0)
2253                 epqstate->es_param_exec_vals = (ParamExecData *)
2254                         palloc0(estate->es_topPlan->nParamExec * sizeof(ParamExecData));
2255         epqstate->es_rowMarks = estate->es_rowMarks;
2256         epqstate->es_instrument = estate->es_instrument;
2257         epqstate->es_select_into = estate->es_select_into;
2258         epqstate->es_into_oids = estate->es_into_oids;
2259         epqstate->es_topPlan = estate->es_topPlan;
2260
2261         /*
2262          * Each epqstate must have its own es_evTupleNull state, but all the stack
2263          * entries share es_evTuple state.      This allows sub-rechecks to inherit
2264          * the value being examined by an outer recheck.
2265          */
2266         epqstate->es_evTupleNull = (bool *) palloc0(rtsize * sizeof(bool));
2267         if (priorepq == NULL)
2268                 /* first PQ stack entry */
2269                 epqstate->es_evTuple = (HeapTuple *)
2270                         palloc0(rtsize * sizeof(HeapTuple));
2271         else
2272                 /* later stack entries share the same storage */
2273                 epqstate->es_evTuple = priorepq->estate->es_evTuple;
2274
2275         epqstate->es_tupleTable =
2276                 ExecCreateTupleTable(estate->es_tupleTable->size);
2277
2278         epq->planstate = ExecInitNode(estate->es_topPlan, epqstate, 0);
2279
2280         MemoryContextSwitchTo(oldcontext);
2281 }
2282
2283 /*
2284  * End execution of one level of PlanQual.
2285  *
2286  * This is a cut-down version of ExecutorEnd(); basically we want to do most
2287  * of the normal cleanup, but *not* close result relations (which we are
2288  * just sharing from the outer query).
2289  */
2290 static void
2291 EvalPlanQualStop(evalPlanQual *epq)
2292 {
2293         EState     *epqstate = epq->estate;
2294         MemoryContext oldcontext;
2295
2296         oldcontext = MemoryContextSwitchTo(epqstate->es_query_cxt);
2297
2298         ExecEndNode(epq->planstate);
2299
2300         ExecDropTupleTable(epqstate->es_tupleTable, true);
2301         epqstate->es_tupleTable = NULL;
2302
2303         if (epqstate->es_evTuple[epq->rti - 1] != NULL)
2304         {
2305                 heap_freetuple(epqstate->es_evTuple[epq->rti - 1]);
2306                 epqstate->es_evTuple[epq->rti - 1] = NULL;
2307         }
2308
2309         MemoryContextSwitchTo(oldcontext);
2310
2311         FreeExecutorState(epqstate);
2312
2313         epq->estate = NULL;
2314         epq->planstate = NULL;
2315 }