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  11  *      The old ExecutorMain() has been replaced by ExecutorStart(),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) 1996-2002, PostgreSQL Global Development Group
  25  * Portions Copyright (c) 1994, Regents of the University of California
  26  *
  27  *
  28  * IDENTIFICATION
  29  *        $Header: /cvsroot/pgsql/src/backend/executor/execMain.c,v 1.204 2003/03/27 14:33:11 tgl Exp $
  30  *
  31  *-------------------------------------------------------------------------
  32  */
  33 #include "postgres.h"
  34
  35 #include "access/heapam.h"
  36 #include "catalog/heap.h"
  37 #include "catalog/namespace.h"
  38 #include "commands/tablecmds.h"
  39 #include "commands/trigger.h"
  40 #include "executor/execdebug.h"
  41 #include "executor/execdefs.h"
  42 #include "miscadmin.h"
  43 #include "optimizer/var.h"
  44 #include "parser/parsetree.h"
  45 #include "utils/acl.h"
  46 #include "utils/lsyscache.h"
  47
  48
  49 typedef struct execRowMark
  50 {
  51         Relation        relation;
  52         Index           rti;
  53         char            resname[32];
  54 } execRowMark;
  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);
  67 static void initResultRelInfo(ResultRelInfo *resultRelInfo,
  68                                   Index resultRelationIndex,
  69                                   List *rangeTable,
  70                                   CmdType operation);
  71 static TupleTableSlot *ExecutePlan(EState *estate, PlanState *planstate,
  72                         CmdType operation,
  73                         long numberTuples,
  74                         ScanDirection direction,
  75                         DestReceiver *destfunc);
  76 static void ExecSelect(TupleTableSlot *slot,
  77                    DestReceiver *destfunc,
  78                    EState *estate);
  79 static void ExecInsert(TupleTableSlot *slot, ItemPointer tupleid,
  80                    EState *estate);
  81 static void ExecDelete(TupleTableSlot *slot, ItemPointer tupleid,
  82                    EState *estate);
  83 static void ExecUpdate(TupleTableSlot *slot, ItemPointer tupleid,
  84                    EState *estate);
  85 static TupleTableSlot *EvalPlanQualNext(EState *estate);
  86 static void EndEvalPlanQual(EState *estate);
  87 static void ExecCheckRTEPerms(RangeTblEntry *rte, CmdType operation);
  88 static void ExecCheckXactReadOnly(Query *parsetree, CmdType operation);
  89 static void EvalPlanQualStart(evalPlanQual *epq, EState *estate,
  90                                                           evalPlanQual *priorepq);
  91 static void EvalPlanQualStop(evalPlanQual *epq);
  92
  93 /* end of local decls */
  94
  95
  96 /* ----------------------------------------------------------------
  97  *              ExecutorStart
  98  *
  99  *              This routine must be called at the beginning of any execution of any
 100  *              query plan
 101  *
 102  * Takes a QueryDesc previously created by CreateQueryDesc (it's not real
 103  * clear why we bother to separate the two functions, but...).  The tupDesc
 104  * field of the QueryDesc is filled in to describe the tuples that will be
 105  * returned, and the internal fields (estate and planstate) are set up.
 106  *
 107  * NB: the CurrentMemoryContext when this is called will become the parent
 108  * of the per-query context used for this Executor invocation.
 109  * ----------------------------------------------------------------
 110  */
 111 void
 112 ExecutorStart(QueryDesc *queryDesc)
 113 {
 114         EState     *estate;
 115         MemoryContext oldcontext;
 116
 117         /* sanity checks: queryDesc must not be started already */
 118         Assert(queryDesc != NULL);
 119         Assert(queryDesc->estate == NULL);
 120
 121         /*
 122          * Build EState, switch into per-query memory context for startup.
 123          */
 124         estate = CreateExecutorState();
 125         queryDesc->estate = estate;
 126
 127         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 128
 129         /*
 130          * Fill in parameters, if any, from queryDesc
 131          */
 132         estate->es_param_list_info = queryDesc->params;
 133
 134         if (queryDesc->plantree->nParamExec > 0)
 135                 estate->es_param_exec_vals = (ParamExecData *)
 136                         palloc0(queryDesc->plantree->nParamExec * sizeof(ParamExecData));
 137
 138         estate->es_instrument = queryDesc->doInstrument;
 139
 140         /*
 141          * Make our own private copy of the current query snapshot data.
 142          *
 143          * This "freezes" our idea of which tuples are good and which are not for
 144          * the life of this query, even if it outlives the current command and
 145          * current snapshot.
 146          */
 147         estate->es_snapshot = CopyQuerySnapshot();
 148
 149         /*
 150          * Initialize the plan state tree
 151          */
 152         InitPlan(queryDesc);
 153
 154         MemoryContextSwitchTo(oldcontext);
 155 }
 156
 157 /* ----------------------------------------------------------------
 158  *              ExecutorRun
 159  *
 160  *              This is the main routine of the executor module. It accepts
 161  *              the query descriptor from the traffic cop and executes the
 162  *              query plan.
 163  *
 164  *              ExecutorStart must have been called already.
 165  *
 166  *              If direction is NoMovementScanDirection then nothing is done
 167  *              except to start up/shut down the destination.  Otherwise,
 168  *              we retrieve up to 'count' tuples in the specified direction.
 169  *
 170  *              Note: count = 0 is interpreted as no portal limit, i.e., run to
 171  *              completion.
 172  *
 173  * ----------------------------------------------------------------
 174  */
 175 TupleTableSlot *
 176 ExecutorRun(QueryDesc *queryDesc,
 177                         ScanDirection direction, long count)
 178 {
 179         EState     *estate;
 180         CmdType         operation;
 181         CommandDest dest;
 182         DestReceiver *destfunc;
 183         TupleTableSlot *result;
 184         MemoryContext oldcontext;
 185
 186         /* sanity checks */
 187         Assert(queryDesc != NULL);
 188
 189         estate = queryDesc->estate;
 190
 191         Assert(estate != NULL);
 192
 193         /*
 194          * Switch into per-query memory context
 195          */
 196         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 197
 198         /*
 199          * extract information from the query descriptor and the query
 200          * feature.
 201          */
 202         operation = queryDesc->operation;
 203         dest = queryDesc->dest;
 204
 205         /*
 206          * If the transaction is read-only, we need to check if any writes
 207          * are planned to non-temporary tables.  This is done here at this
 208          * rather late stage so that we can handle EXPLAIN vs. EXPLAIN
 209          * ANALYZE easily.
 210          */
 211         ExecCheckXactReadOnly(queryDesc->parsetree, operation);
 212
 213         /*
 214          * startup tuple receiver
 215          */
 216         estate->es_processed = 0;
 217         estate->es_lastoid = InvalidOid;
 218
 219         destfunc = DestToFunction(dest);
 220         (*destfunc->setup) (destfunc, (int) operation,
 221                                                 queryDesc->portalName, queryDesc->tupDesc);
 222
 223         /*
 224          * run plan
 225          */
 226         if (direction == NoMovementScanDirection)
 227                 result = NULL;
 228         else
 229                 result = ExecutePlan(estate,
 230                                                          queryDesc->planstate,
 231                                                          operation,
 232                                                          count,
 233                                                          direction,
 234                                                          destfunc);
 235
 236         /*
 237          * shutdown receiver
 238          */
 239         (*destfunc->cleanup) (destfunc);
 240
 241         MemoryContextSwitchTo(oldcontext);
 242
 243         return result;
 244 }
 245
 246 /* ----------------------------------------------------------------
 247  *              ExecutorEnd
 248  *
 249  *              This routine must be called at the end of execution of any
 250  *              query plan
 251  * ----------------------------------------------------------------
 252  */
 253 void
 254 ExecutorEnd(QueryDesc *queryDesc)
 255 {
 256         EState     *estate;
 257         MemoryContext oldcontext;
 258
 259         /* sanity checks */
 260         Assert(queryDesc != NULL);
 261
 262         estate = queryDesc->estate;
 263
 264         Assert(estate != NULL);
 265
 266         /*
 267          * Switch into per-query memory context to run ExecEndPlan
 268          */
 269         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 270
 271         ExecEndPlan(queryDesc->planstate, estate);
 272
 273         /*
 274          * Must switch out of context before destroying it
 275          */
 276         MemoryContextSwitchTo(oldcontext);
 277
 278         /*
 279          * Release EState and per-query memory context.  This should release
 280          * everything the executor has allocated.
 281          */
 282         FreeExecutorState(estate);
 283
 284         /* Reset queryDesc fields that no longer point to anything */
 285         queryDesc->tupDesc = NULL;
 286         queryDesc->estate = NULL;
 287         queryDesc->planstate = NULL;
 288 }
 289
 290 /* ----------------------------------------------------------------
 291  *              ExecutorRewind
 292  *
 293  *              This routine may be called on an open queryDesc to rewind it
 294  *              to the start.
 295  * ----------------------------------------------------------------
 296  */
 297 void
 298 ExecutorRewind(QueryDesc *queryDesc)
 299 {
 300         EState     *estate;
 301         MemoryContext oldcontext;
 302
 303         /* sanity checks */
 304         Assert(queryDesc != NULL);
 305
 306         estate = queryDesc->estate;
 307
 308         Assert(estate != NULL);
 309
 310         /* It's probably not sensible to rescan updating queries */
 311         Assert(queryDesc->operation == CMD_SELECT);
 312
 313         /*
 314          * Switch into per-query memory context
 315          */
 316         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 317
 318         /*
 319          * rescan plan
 320          */
 321         ExecReScan(queryDesc->planstate, NULL);
 322
 323         MemoryContextSwitchTo(oldcontext);
 324 }
 325
 326
 327 /*
 328  * ExecCheckRTPerms
 329  *              Check access permissions for all relations listed in a range table.
 330  */
 331 void
 332 ExecCheckRTPerms(List *rangeTable, CmdType operation)
 333 {
 334         List       *lp;
 335
 336         foreach(lp, rangeTable)
 337         {
 338                 RangeTblEntry *rte = lfirst(lp);
 339
 340                 ExecCheckRTEPerms(rte, operation);
 341         }
 342 }
 343
 344 /*
 345  * ExecCheckRTEPerms
 346  *              Check access permissions for a single RTE.
 347  */
 348 static void
 349 ExecCheckRTEPerms(RangeTblEntry *rte, CmdType operation)
 350 {
 351         Oid                     relOid;
 352         AclId           userid;
 353         AclResult       aclcheck_result;
 354
 355         /*
 356          * If it's a subquery, recursively examine its rangetable.
 357          */
 358         if (rte->rtekind == RTE_SUBQUERY)
 359         {
 360                 ExecCheckRTPerms(rte->subquery->rtable, operation);
 361                 return;
 362         }
 363
 364         /*
 365          * Otherwise, only plain-relation RTEs need to be checked here.
 366          * Function RTEs are checked by init_fcache when the function is prepared
 367          * for execution. Join and special RTEs need no checks.
 368          */
 369         if (rte->rtekind != RTE_RELATION)
 370                 return;
 371
 372         relOid = rte->relid;
 373
 374         /*
 375          * userid to check as: current user unless we have a setuid
 376          * indication.
 377          *
 378          * Note: GetUserId() is presently fast enough that there's no harm in
 379          * calling it separately for each RTE.  If that stops being true, we
 380          * could call it once in ExecCheckRTPerms and pass the userid down
 381          * from there.  But for now, no need for the extra clutter.
 382          */
 383         userid = rte->checkAsUser ? rte->checkAsUser : GetUserId();
 384
 385 #define CHECK(MODE)             pg_class_aclcheck(relOid, userid, MODE)
 386
 387         if (rte->checkForRead)
 388         {
 389                 aclcheck_result = CHECK(ACL_SELECT);
 390                 if (aclcheck_result != ACLCHECK_OK)
 391                         aclcheck_error(aclcheck_result, get_rel_name(relOid));
 392         }
 393
 394         if (rte->checkForWrite)
 395         {
 396                 /*
 397                  * Note: write access in a SELECT context means SELECT FOR UPDATE.
 398                  * Right now we don't distinguish that from true update as far as
 399                  * permissions checks are concerned.
 400                  */
 401                 switch (operation)
 402                 {
 403                         case CMD_INSERT:
 404                                 aclcheck_result = CHECK(ACL_INSERT);
 405                                 break;
 406                         case CMD_SELECT:
 407                         case CMD_UPDATE:
 408                                 aclcheck_result = CHECK(ACL_UPDATE);
 409                                 break;
 410                         case CMD_DELETE:
 411                                 aclcheck_result = CHECK(ACL_DELETE);
 412                                 break;
 413                         default:
 414                                 elog(ERROR, "ExecCheckRTEPerms: bogus operation %d",
 415                                          operation);
 416                                 aclcheck_result = ACLCHECK_OK;  /* keep compiler quiet */
 417                                 break;
 418                 }
 419                 if (aclcheck_result != ACLCHECK_OK)
 420                         aclcheck_error(aclcheck_result, get_rel_name(relOid));
 421         }
 422 }
 423
 424
 425 /* ===============================================================
 426  * ===============================================================
 427                                                  static routines follow
 428  * ===============================================================
 429  * ===============================================================
 430  */
 431
 432
 433 static void
 434 ExecCheckXactReadOnly(Query *parsetree, CmdType operation)
 435 {
 436         if (!XactReadOnly)
 437                 return;
 438
 439         /* CREATE TABLE AS or SELECT INTO */
 440         if (operation == CMD_SELECT && parsetree->into != NULL)
 441                 goto fail;
 442
 443         if (operation == CMD_DELETE || operation == CMD_INSERT
 444                 || operation == CMD_UPDATE)
 445         {
 446                 List *lp;
 447
 448                 foreach(lp, parsetree->rtable)
 449                 {
 450                         RangeTblEntry *rte = lfirst(lp);
 451
 452                         if (rte->rtekind != RTE_RELATION)
 453                                 continue;
 454
 455                         if (!rte->checkForWrite)
 456                                 continue;
 457
 458                         if (isTempNamespace(get_rel_namespace(rte->relid)))
 459                                 continue;
 460
 461                         goto fail;
 462                 }
 463         }
 464
 465         return;
 466
 467 fail:
 468         elog(ERROR, "transaction is read-only");
 469 }
 470
 471
 472 /* ----------------------------------------------------------------
 473  *              InitPlan
 474  *
 475  *              Initializes the query plan: open files, allocate storage
 476  *              and start up the rule manager
 477  * ----------------------------------------------------------------
 478  */
 479 static void
 480 InitPlan(QueryDesc *queryDesc)
 481 {
 482         CmdType         operation = queryDesc->operation;
 483         Query *parseTree = queryDesc->parsetree;
 484         Plan *plan = queryDesc->plantree;
 485         EState *estate = queryDesc->estate;
 486         PlanState  *planstate;
 487         List       *rangeTable;
 488         Relation        intoRelationDesc;
 489         bool            do_select_into;
 490         TupleDesc       tupType;
 491
 492         /*
 493          * Do permissions checks.  It's sufficient to examine the query's
 494          * top rangetable here --- subplan RTEs will be checked during
 495          * ExecInitSubPlan().
 496          */
 497         ExecCheckRTPerms(parseTree->rtable, operation);
 498
 499         /*
 500          * get information from query descriptor
 501          */
 502         rangeTable = parseTree->rtable;
 503
 504         /*
 505          * initialize the node's execution state
 506          */
 507         estate->es_range_table = rangeTable;
 508
 509         /*
 510          * if there is a result relation, initialize result relation stuff
 511          */
 512         if (parseTree->resultRelation != 0 && operation != CMD_SELECT)
 513         {
 514                 List       *resultRelations = parseTree->resultRelations;
 515                 int                     numResultRelations;
 516                 ResultRelInfo *resultRelInfos;
 517
 518                 if (resultRelations != NIL)
 519                 {
 520                         /*
 521                          * Multiple result relations (due to inheritance)
 522                          * parseTree->resultRelations identifies them all
 523                          */
 524                         ResultRelInfo *resultRelInfo;
 525
 526                         numResultRelations = length(resultRelations);
 527                         resultRelInfos = (ResultRelInfo *)
 528                                 palloc(numResultRelations * sizeof(ResultRelInfo));
 529                         resultRelInfo = resultRelInfos;
 530                         while (resultRelations != NIL)
 531                         {
 532                                 initResultRelInfo(resultRelInfo,
 533                                                                   lfirsti(resultRelations),
 534                                                                   rangeTable,
 535                                                                   operation);
 536                                 resultRelInfo++;
 537                                 resultRelations = lnext(resultRelations);
 538                         }
 539                 }
 540                 else
 541                 {
 542                         /*
 543                          * Single result relation identified by
 544                          * parseTree->resultRelation
 545                          */
 546                         numResultRelations = 1;
 547                         resultRelInfos = (ResultRelInfo *) palloc(sizeof(ResultRelInfo));
 548                         initResultRelInfo(resultRelInfos,
 549                                                           parseTree->resultRelation,
 550                                                           rangeTable,
 551                                                           operation);
 552                 }
 553
 554                 estate->es_result_relations = resultRelInfos;
 555                 estate->es_num_result_relations = numResultRelations;
 556                 /* Initialize to first or only result rel */
 557                 estate->es_result_relation_info = resultRelInfos;
 558         }
 559         else
 560         {
 561                 /*
 562                  * if no result relation, then set state appropriately
 563                  */
 564                 estate->es_result_relations = NULL;
 565                 estate->es_num_result_relations = 0;
 566                 estate->es_result_relation_info = NULL;
 567         }
 568
 569         /*
 570          * Detect whether we're doing SELECT INTO.  If so, set the force_oids
 571          * flag appropriately so that the plan tree will be initialized with
 572          * the correct tuple descriptors.
 573          */
 574         do_select_into = false;
 575
 576         if (operation == CMD_SELECT && parseTree->into != NULL)
 577         {
 578                 do_select_into = true;
 579                 /*
 580                  * For now, always create OIDs in SELECT INTO; this is for backwards
 581                  * compatibility with pre-7.3 behavior.  Eventually we might want
 582                  * to allow the user to choose.
 583                  */
 584                 estate->es_force_oids = true;
 585         }
 586
 587         /*
 588          * Have to lock relations selected for update
 589          */
 590         estate->es_rowMark = NIL;
 591         if (parseTree->rowMarks != NIL)
 592         {
 593                 List       *l;
 594
 595                 foreach(l, parseTree->rowMarks)
 596                 {
 597                         Index           rti = lfirsti(l);
 598                         Oid                     relid = getrelid(rti, rangeTable);
 599                         Relation        relation;
 600                         execRowMark *erm;
 601
 602                         relation = heap_open(relid, RowShareLock);
 603                         erm = (execRowMark *) palloc(sizeof(execRowMark));
 604                         erm->relation = relation;
 605                         erm->rti = rti;
 606                         snprintf(erm->resname, sizeof(erm->resname), "ctid%u", rti);
 607                         estate->es_rowMark = lappend(estate->es_rowMark, erm);
 608                 }
 609         }
 610
 611         /*
 612          * initialize the executor "tuple" table.  We need slots for all the
 613          * plan nodes, plus possibly output slots for the junkfilter(s). At
 614          * this point we aren't sure if we need junkfilters, so just add slots
 615          * for them unconditionally.
 616          */
 617         {
 618                 int                     nSlots = ExecCountSlotsNode(plan);
 619
 620                 if (parseTree->resultRelations != NIL)
 621                         nSlots += length(parseTree->resultRelations);
 622                 else
 623                         nSlots += 1;
 624                 estate->es_tupleTable = ExecCreateTupleTable(nSlots);
 625         }
 626
 627         /* mark EvalPlanQual not active */
 628         estate->es_topPlan = plan;
 629         estate->es_evalPlanQual = NULL;
 630         estate->es_evTupleNull = NULL;
 631         estate->es_evTuple = NULL;
 632         estate->es_useEvalPlan = false;
 633
 634         /*
 635          * initialize the private state information for all the nodes in the
 636          * query tree.  This opens files, allocates storage and leaves us
 637          * ready to start processing tuples.
 638          */
 639         planstate = ExecInitNode(plan, estate);
 640
 641         /*
 642          * Get the tuple descriptor describing the type of tuples to return.
 643          * (this is especially important if we are creating a relation with
 644          * "SELECT INTO")
 645          */
 646         tupType = ExecGetTupType(planstate);
 647
 648         /*
 649          * Initialize the junk filter if needed.  SELECT and INSERT queries need a
 650          * filter if there are any junk attrs in the tlist.  INSERT and SELECT
 651          * INTO also need a filter if the top plan node is a scan node that's not
 652          * doing projection (else we'll be scribbling on the scan tuple!)  UPDATE
 653          * and DELETE always need a filter, since there's always a junk 'ctid'
 654          * attribute present --- no need to look first.
 655          */
 656         {
 657                 bool            junk_filter_needed = false;
 658                 List       *tlist;
 659
 660                 switch (operation)
 661                 {
 662                         case CMD_SELECT:
 663                         case CMD_INSERT:
 664                                 foreach(tlist, plan->targetlist)
 665                                 {
 666                                         TargetEntry *tle = (TargetEntry *) lfirst(tlist);
 667
 668                                         if (tle->resdom->resjunk)
 669                                         {
 670                                                 junk_filter_needed = true;
 671                                                 break;
 672                                         }
 673                                 }
 674                                 if (!junk_filter_needed &&
 675                                         (operation == CMD_INSERT || do_select_into))
 676                                 {
 677                                         if (IsA(planstate, SeqScanState) ||
 678                                                 IsA(planstate, IndexScanState) ||
 679                                                 IsA(planstate, TidScanState) ||
 680                                                 IsA(planstate, SubqueryScanState) ||
 681                                                 IsA(planstate, FunctionScanState))
 682                                         {
 683                                                 if (planstate->ps_ProjInfo == NULL)
 684                                                         junk_filter_needed = true;
 685                                         }
 686                                 }
 687                                 break;
 688                         case CMD_UPDATE:
 689                         case CMD_DELETE:
 690                                 junk_filter_needed = true;
 691                                 break;
 692                         default:
 693                                 break;
 694                 }
 695
 696                 if (junk_filter_needed)
 697                 {
 698                         /*
 699                          * If there are multiple result relations, each one needs its
 700                          * own junk filter.  Note this is only possible for
 701                          * UPDATE/DELETE, so we can't be fooled by some needing a
 702                          * filter and some not.
 703                          */
 704                         if (parseTree->resultRelations != NIL)
 705                         {
 706                                 PlanState **appendplans;
 707                                 int                     as_nplans;
 708                                 ResultRelInfo *resultRelInfo;
 709                                 int                     i;
 710
 711                                 /* Top plan had better be an Append here. */
 712                                 Assert(IsA(plan, Append));
 713                                 Assert(((Append *) plan)->isTarget);
 714                                 Assert(IsA(planstate, AppendState));
 715                                 appendplans = ((AppendState *) planstate)->appendplans;
 716                                 as_nplans = ((AppendState *) planstate)->as_nplans;
 717                                 Assert(as_nplans == estate->es_num_result_relations);
 718                                 resultRelInfo = estate->es_result_relations;
 719                                 for (i = 0; i < as_nplans; i++)
 720                                 {
 721                                         PlanState  *subplan = appendplans[i];
 722                                         JunkFilter *j;
 723
 724                                         j = ExecInitJunkFilter(subplan->plan->targetlist,
 725                                                                                    ExecGetTupType(subplan),
 726                                                           ExecAllocTableSlot(estate->es_tupleTable));
 727                                         resultRelInfo->ri_junkFilter = j;
 728                                         resultRelInfo++;
 729                                 }
 730
 731                                 /*
 732                                  * Set active junkfilter too; at this point ExecInitAppend
 733                                  * has already selected an active result relation...
 734                                  */
 735                                 estate->es_junkFilter =
 736                                         estate->es_result_relation_info->ri_junkFilter;
 737                         }
 738                         else
 739                         {
 740                                 /* Normal case with just one JunkFilter */
 741                                 JunkFilter *j;
 742
 743                                 j = ExecInitJunkFilter(planstate->plan->targetlist,
 744                                                                            tupType,
 745                                                           ExecAllocTableSlot(estate->es_tupleTable));
 746                                 estate->es_junkFilter = j;
 747                                 if (estate->es_result_relation_info)
 748                                         estate->es_result_relation_info->ri_junkFilter = j;
 749
 750                                 /* For SELECT, want to return the cleaned tuple type */
 751                                 if (operation == CMD_SELECT)
 752                                         tupType = j->jf_cleanTupType;
 753                         }
 754                 }
 755                 else
 756                         estate->es_junkFilter = NULL;
 757         }
 758
 759         /*
 760          * If doing SELECT INTO, initialize the "into" relation.  We must wait
 761          * till now so we have the "clean" result tuple type to create the
 762          * new table from.
 763          */
 764         intoRelationDesc = (Relation) NULL;
 765
 766         if (do_select_into)
 767         {
 768                 char       *intoName;
 769                 Oid                     namespaceId;
 770                 AclResult       aclresult;
 771                 Oid                     intoRelationId;
 772                 TupleDesc       tupdesc;
 773
 774                 /*
 775                  * find namespace to create in, check permissions
 776                  */
 777                 intoName = parseTree->into->relname;
 778                 namespaceId = RangeVarGetCreationNamespace(parseTree->into);
 779
 780                 aclresult = pg_namespace_aclcheck(namespaceId, GetUserId(),
 781                                                                                   ACL_CREATE);
 782                 if (aclresult != ACLCHECK_OK)
 783                         aclcheck_error(aclresult, get_namespace_name(namespaceId));
 784
 785                 /*
 786                  * have to copy tupType to get rid of constraints
 787                  */
 788                 tupdesc = CreateTupleDescCopy(tupType);
 789
 790                 intoRelationId = heap_create_with_catalog(intoName,
 791                                                                                                   namespaceId,
 792                                                                                                   tupdesc,
 793                                                                                                   RELKIND_RELATION,
 794                                                                                                   false,
 795                                                                                                   ONCOMMIT_NOOP,
 796                                                                                                   allowSystemTableMods);
 797
 798                 FreeTupleDesc(tupdesc);
 799
 800                 /*
 801                  * Advance command counter so that the newly-created
 802                  * relation's catalog tuples will be visible to heap_open.
 803                  */
 804                 CommandCounterIncrement();
 805
 806                 /*
 807                  * If necessary, create a TOAST table for the into
 808                  * relation. Note that AlterTableCreateToastTable ends
 809                  * with CommandCounterIncrement(), so that the TOAST table
 810                  * will be visible for insertion.
 811                  */
 812                 AlterTableCreateToastTable(intoRelationId, true);
 813
 814                 /*
 815                  * And open the constructed table for writing.
 816                  */
 817                 intoRelationDesc = heap_open(intoRelationId, AccessExclusiveLock);
 818         }
 819
 820         estate->es_into_relation_descriptor = intoRelationDesc;
 821
 822         queryDesc->tupDesc = tupType;
 823         queryDesc->planstate = planstate;
 824 }
 825
 826 /*
 827  * Initialize ResultRelInfo data for one result relation
 828  */
 829 static void
 830 initResultRelInfo(ResultRelInfo *resultRelInfo,
 831                                   Index resultRelationIndex,
 832                                   List *rangeTable,
 833                                   CmdType operation)
 834 {
 835         Oid                     resultRelationOid;
 836         Relation        resultRelationDesc;
 837
 838         resultRelationOid = getrelid(resultRelationIndex, rangeTable);
 839         resultRelationDesc = heap_open(resultRelationOid, RowExclusiveLock);
 840
 841         switch (resultRelationDesc->rd_rel->relkind)
 842         {
 843                 case RELKIND_SEQUENCE:
 844                         elog(ERROR, "You can't change sequence relation %s",
 845                                  RelationGetRelationName(resultRelationDesc));
 846                         break;
 847                 case RELKIND_TOASTVALUE:
 848                         elog(ERROR, "You can't change toast relation %s",
 849                                  RelationGetRelationName(resultRelationDesc));
 850                         break;
 851                 case RELKIND_VIEW:
 852                         elog(ERROR, "You can't change view relation %s",
 853                                  RelationGetRelationName(resultRelationDesc));
 854                         break;
 855         }
 856
 857         MemSet(resultRelInfo, 0, sizeof(ResultRelInfo));
 858         resultRelInfo->type = T_ResultRelInfo;
 859         resultRelInfo->ri_RangeTableIndex = resultRelationIndex;
 860         resultRelInfo->ri_RelationDesc = resultRelationDesc;
 861         resultRelInfo->ri_NumIndices = 0;
 862         resultRelInfo->ri_IndexRelationDescs = NULL;
 863         resultRelInfo->ri_IndexRelationInfo = NULL;
 864         /* make a copy so as not to depend on relcache info not changing... */
 865         resultRelInfo->ri_TrigDesc = CopyTriggerDesc(resultRelationDesc->trigdesc);
 866         resultRelInfo->ri_TrigFunctions = NULL;
 867         resultRelInfo->ri_ConstraintExprs = NULL;
 868         resultRelInfo->ri_junkFilter = NULL;
 869
 870         /*
 871          * If there are indices on the result relation, open them and save
 872          * descriptors in the result relation info, so that we can add new
 873          * index entries for the tuples we add/update.  We need not do this
 874          * for a DELETE, however, since deletion doesn't affect indexes.
 875          */
 876         if (resultRelationDesc->rd_rel->relhasindex &&
 877                 operation != CMD_DELETE)
 878                 ExecOpenIndices(resultRelInfo);
 879 }
 880
 881 /* ----------------------------------------------------------------
 882  *              ExecEndPlan
 883  *
 884  *              Cleans up the query plan -- closes files and frees up storage
 885  *
 886  * NOTE: we are no longer very worried about freeing storage per se
 887  * in this code; FreeExecutorState should be guaranteed to release all
 888  * memory that needs to be released.  What we are worried about doing
 889  * is closing relations and dropping buffer pins.  Thus, for example,
 890  * tuple tables must be cleared or dropped to ensure pins are released.
 891  * ----------------------------------------------------------------
 892  */
 893 void
 894 ExecEndPlan(PlanState *planstate, EState *estate)
 895 {
 896         ResultRelInfo *resultRelInfo;
 897         int                     i;
 898         List       *l;
 899
 900         /*
 901          * shut down any PlanQual processing we were doing
 902          */
 903         if (estate->es_evalPlanQual != NULL)
 904                 EndEvalPlanQual(estate);
 905
 906         /*
 907          * shut down the node-type-specific query processing
 908          */
 909         ExecEndNode(planstate);
 910
 911         /*
 912          * destroy the executor "tuple" table.
 913          */
 914         ExecDropTupleTable(estate->es_tupleTable, true);
 915         estate->es_tupleTable = NULL;
 916
 917         /*
 918          * close the result relation(s) if any, but hold locks until xact
 919          * commit.
 920          */
 921         resultRelInfo = estate->es_result_relations;
 922         for (i = estate->es_num_result_relations; i > 0; i--)
 923         {
 924                 /* Close indices and then the relation itself */
 925                 ExecCloseIndices(resultRelInfo);
 926                 heap_close(resultRelInfo->ri_RelationDesc, NoLock);
 927                 resultRelInfo++;
 928         }
 929
 930         /*
 931          * close the "into" relation if necessary, again keeping lock
 932          */
 933         if (estate->es_into_relation_descriptor != NULL)
 934                 heap_close(estate->es_into_relation_descriptor, NoLock);
 935
 936         /*
 937          * close any relations selected FOR UPDATE, again keeping locks
 938          */
 939         foreach(l, estate->es_rowMark)
 940         {
 941                 execRowMark *erm = lfirst(l);
 942
 943                 heap_close(erm->relation, NoLock);
 944         }
 945 }
 946
 947 /* ----------------------------------------------------------------
 948  *              ExecutePlan
 949  *
 950  *              processes the query plan to retrieve 'numberTuples' tuples in the
 951  *              direction specified.
 952  *
 953  *              Retrieves all tuples if numberTuples is 0
 954  *
 955  *              result is either a slot containing the last tuple in the case
 956  *              of a SELECT or NULL otherwise.
 957  *
 958  * Note: the ctid attribute is a 'junk' attribute that is removed before the
 959  * user can see it
 960  * ----------------------------------------------------------------
 961  */
 962 static TupleTableSlot *
 963 ExecutePlan(EState *estate,
 964                         PlanState *planstate,
 965                         CmdType operation,
 966                         long numberTuples,
 967                         ScanDirection direction,
 968                         DestReceiver *destfunc)
 969 {
 970         JunkFilter                      *junkfilter;
 971         TupleTableSlot          *slot;
 972         ItemPointer                      tupleid = NULL;
 973         ItemPointerData          tuple_ctid;
 974         long                             current_tuple_count;
 975         TupleTableSlot          *result;
 976
 977         /*
 978          * initialize local variables
 979          */
 980         slot = NULL;
 981         current_tuple_count = 0;
 982         result = NULL;
 983
 984         /*
 985          * Set the direction.
 986          */
 987         estate->es_direction = direction;
 988
 989         /*
 990          * Process BEFORE EACH STATEMENT triggers
 991          */
 992         switch (operation)
 993         {
 994                 case CMD_UPDATE:
 995                         ExecBSUpdateTriggers(estate, estate->es_result_relation_info);
 996                         break;
 997                 case CMD_DELETE:
 998                         ExecBSDeleteTriggers(estate, estate->es_result_relation_info);
 999                         break;
1000                 case CMD_INSERT:
1001                         ExecBSInsertTriggers(estate, estate->es_result_relation_info);
1002                         break;
1003                 default:
1004                         /* do nothing */
1005                         break;
1006         }
1007
1008         /*
1009          * Loop until we've processed the proper number of tuples from the
1010          * plan.
1011          */
1012
1013         for (;;)
1014         {
1015                 /* Reset the per-output-tuple exprcontext */
1016                 ResetPerTupleExprContext(estate);
1017
1018                 /*
1019                  * Execute the plan and obtain a tuple
1020                  */
1021 lnext:  ;
1022                 if (estate->es_useEvalPlan)
1023                 {
1024                         slot = EvalPlanQualNext(estate);
1025                         if (TupIsNull(slot))
1026                                 slot = ExecProcNode(planstate);
1027                 }
1028                 else
1029                         slot = ExecProcNode(planstate);
1030
1031                 /*
1032                  * if the tuple is null, then we assume there is nothing more to
1033                  * process so we just return null...
1034                  */
1035                 if (TupIsNull(slot))
1036                 {
1037                         result = NULL;
1038                         break;
1039                 }
1040
1041                 /*
1042                  * if we have a junk filter, then project a new tuple with the
1043                  * junk removed.
1044                  *
1045                  * Store this new "clean" tuple in the junkfilter's resultSlot.
1046                  * (Formerly, we stored it back over the "dirty" tuple, which is
1047                  * WRONG because that tuple slot has the wrong descriptor.)
1048                  *
1049                  * Also, extract all the junk information we need.
1050                  */
1051                 if ((junkfilter = estate->es_junkFilter) != (JunkFilter *) NULL)
1052                 {
1053                         Datum           datum;
1054                         HeapTuple       newTuple;
1055                         bool            isNull;
1056
1057                         /*
1058                          * extract the 'ctid' junk attribute.
1059                          */
1060                         if (operation == CMD_UPDATE || operation == CMD_DELETE)
1061                         {
1062                                 if (!ExecGetJunkAttribute(junkfilter,
1063                                                                                   slot,
1064                                                                                   "ctid",
1065                                                                                   &datum,
1066                                                                                   &isNull))
1067                                         elog(ERROR, "ExecutePlan: NO (junk) `ctid' was found!");
1068
1069                                 /* shouldn't ever get a null result... */
1070                                 if (isNull)
1071                                         elog(ERROR, "ExecutePlan: (junk) `ctid' is NULL!");
1072
1073                                 tupleid = (ItemPointer) DatumGetPointer(datum);
1074                                 tuple_ctid = *tupleid;  /* make sure we don't free the
1075                                                                                  * ctid!! */
1076                                 tupleid = &tuple_ctid;
1077                         }
1078                         else if (estate->es_rowMark != NIL)
1079                         {
1080                                 List       *l;
1081
1082                 lmark:  ;
1083                                 foreach(l, estate->es_rowMark)
1084                                 {
1085                                         execRowMark *erm = lfirst(l);
1086                                         Buffer          buffer;
1087                                         HeapTupleData tuple;
1088                                         TupleTableSlot *newSlot;
1089                                         int                     test;
1090
1091                                         if (!ExecGetJunkAttribute(junkfilter,
1092                                                                                           slot,
1093                                                                                           erm->resname,
1094                                                                                           &datum,
1095                                                                                           &isNull))
1096                                                 elog(ERROR, "ExecutePlan: NO (junk) `%s' was found!",
1097                                                          erm->resname);
1098
1099                                         /* shouldn't ever get a null result... */
1100                                         if (isNull)
1101                                                 elog(ERROR, "ExecutePlan: (junk) `%s' is NULL!",
1102                                                          erm->resname);
1103
1104                                         tuple.t_self = *((ItemPointer) DatumGetPointer(datum));
1105                                         test = heap_mark4update(erm->relation, &tuple, &buffer,
1106                                                                                         estate->es_snapshot->curcid);
1107                                         ReleaseBuffer(buffer);
1108                                         switch (test)
1109                                         {
1110                                                 case HeapTupleSelfUpdated:
1111                                                         /* treat it as deleted; do not process */
1112                                                         goto lnext;
1113
1114                                                 case HeapTupleMayBeUpdated:
1115                                                         break;
1116
1117                                                 case HeapTupleUpdated:
1118                                                         if (XactIsoLevel == XACT_SERIALIZABLE)
1119                                                                 elog(ERROR, "Can't serialize access due to concurrent update");
1120                                                         if (!(ItemPointerEquals(&(tuple.t_self),
1121                                                                   (ItemPointer) DatumGetPointer(datum))))
1122                                                         {
1123                                                                 newSlot = EvalPlanQual(estate, erm->rti, &(tuple.t_self));
1124                                                                 if (!(TupIsNull(newSlot)))
1125                                                                 {
1126                                                                         slot = newSlot;
1127                                                                         estate->es_useEvalPlan = true;
1128                                                                         goto lmark;
1129                                                                 }
1130                                                         }
1131
1132                                                         /*
1133                                                          * if tuple was deleted or PlanQual failed for
1134                                                          * updated tuple - we must not return this
1135                                                          * tuple!
1136                                                          */
1137                                                         goto lnext;
1138
1139                                                 default:
1140                                                         elog(ERROR, "Unknown status %u from heap_mark4update", test);
1141                                                         return (NULL);
1142                                         }
1143                                 }
1144                         }
1145
1146                         /*
1147                          * Finally create a new "clean" tuple with all junk attributes
1148                          * removed
1149                          */
1150                         newTuple = ExecRemoveJunk(junkfilter, slot);
1151
1152                         slot = ExecStoreTuple(newTuple,         /* tuple to store */
1153                                                                   junkfilter->jf_resultSlot,    /* dest slot */
1154                                                                   InvalidBuffer,                /* this tuple has no
1155                                                                                                                  * buffer */
1156                                                                   true);                /* tuple should be pfreed */
1157                 }
1158
1159                 /*
1160                  * now that we have a tuple, do the appropriate thing with it..
1161                  * either return it to the user, add it to a relation someplace,
1162                  * delete it from a relation, or modify some of its attributes.
1163                  */
1164                 switch (operation)
1165                 {
1166                         case CMD_SELECT:
1167                                 ExecSelect(slot,        /* slot containing tuple */
1168                                                    destfunc,    /* destination's tuple-receiver
1169                                                                                  * obj */
1170                                                    estate);
1171                                 result = slot;
1172                                 break;
1173
1174                         case CMD_INSERT:
1175                                 ExecInsert(slot, tupleid, estate);
1176                                 result = NULL;
1177                                 break;
1178
1179                         case CMD_DELETE:
1180                                 ExecDelete(slot, tupleid, estate);
1181                                 result = NULL;
1182                                 break;
1183
1184                         case CMD_UPDATE:
1185                                 ExecUpdate(slot, tupleid, estate);
1186                                 result = NULL;
1187                                 break;
1188
1189                         default:
1190                                 elog(LOG, "ExecutePlan: unknown operation in queryDesc");
1191                                 result = NULL;
1192                                 break;
1193                 }
1194
1195                 /*
1196                  * check our tuple count.. if we've processed the proper number
1197                  * then quit, else loop again and process more tuples.  Zero
1198                  * numberTuples means no limit.
1199                  */
1200                 current_tuple_count++;
1201                 if (numberTuples && numberTuples == current_tuple_count)
1202                         break;
1203         }
1204
1205         /*
1206          * Process AFTER EACH STATEMENT triggers
1207          */
1208         switch (operation)
1209         {
1210                 case CMD_UPDATE:
1211                         ExecASUpdateTriggers(estate, estate->es_result_relation_info);
1212                         break;
1213                 case CMD_DELETE:
1214                         ExecASDeleteTriggers(estate, estate->es_result_relation_info);
1215                         break;
1216                 case CMD_INSERT:
1217                         ExecASInsertTriggers(estate, estate->es_result_relation_info);
1218                         break;
1219                 default:
1220                         /* do nothing */
1221                         break;
1222         }
1223
1224         /*
1225          * here, result is either a slot containing a tuple in the case of a
1226          * SELECT or NULL otherwise.
1227          */
1228         return result;
1229 }
1230
1231 /* ----------------------------------------------------------------
1232  *              ExecSelect
1233  *
1234  *              SELECTs are easy.. we just pass the tuple to the appropriate
1235  *              print function.  The only complexity is when we do a
1236  *              "SELECT INTO", in which case we insert the tuple into
1237  *              the appropriate relation (note: this is a newly created relation
1238  *              so we don't need to worry about indices or locks.)
1239  * ----------------------------------------------------------------
1240  */
1241 static void
1242 ExecSelect(TupleTableSlot *slot,
1243                    DestReceiver *destfunc,
1244                    EState *estate)
1245 {
1246         HeapTuple       tuple;
1247         TupleDesc       attrtype;
1248
1249         /*
1250          * get the heap tuple out of the tuple table slot
1251          */
1252         tuple = slot->val;
1253         attrtype = slot->ttc_tupleDescriptor;
1254
1255         /*
1256          * insert the tuple into the "into relation"
1257          */
1258         if (estate->es_into_relation_descriptor != NULL)
1259         {
1260                 heap_insert(estate->es_into_relation_descriptor, tuple,
1261                                         estate->es_snapshot->curcid);
1262                 IncrAppended();
1263         }
1264
1265         /*
1266          * send the tuple to the front end (or the screen)
1267          */
1268         (*destfunc->receiveTuple) (tuple, attrtype, destfunc);
1269         IncrRetrieved();
1270         (estate->es_processed)++;
1271 }
1272
1273 /* ----------------------------------------------------------------
1274  *              ExecInsert
1275  *
1276  *              INSERTs are trickier.. we have to insert the tuple into
1277  *              the base relation and insert appropriate tuples into the
1278  *              index relations.
1279  * ----------------------------------------------------------------
1280  */
1281 static void
1282 ExecInsert(TupleTableSlot *slot,
1283                    ItemPointer tupleid,
1284                    EState *estate)
1285 {
1286         HeapTuple       tuple;
1287         ResultRelInfo *resultRelInfo;
1288         Relation        resultRelationDesc;
1289         int                     numIndices;
1290         Oid                     newId;
1291
1292         /*
1293          * get the heap tuple out of the tuple table slot
1294          */
1295         tuple = slot->val;
1296
1297         /*
1298          * get information on the (current) result relation
1299          */
1300         resultRelInfo = estate->es_result_relation_info;
1301         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1302
1303         /* BEFORE ROW INSERT Triggers */
1304         if (resultRelInfo->ri_TrigDesc &&
1305                 resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_INSERT] > 0)
1306         {
1307                 HeapTuple       newtuple;
1308
1309                 newtuple = ExecBRInsertTriggers(estate, resultRelInfo, tuple);
1310
1311                 if (newtuple == NULL)   /* "do nothing" */
1312                         return;
1313
1314                 if (newtuple != tuple)  /* modified by Trigger(s) */
1315                 {
1316                         /*
1317                          * Insert modified tuple into tuple table slot, replacing the
1318                          * original.  We assume that it was allocated in per-tuple
1319                          * memory context, and therefore will go away by itself. The
1320                          * tuple table slot should not try to clear it.
1321                          */
1322                         ExecStoreTuple(newtuple, slot, InvalidBuffer, false);
1323                         tuple = newtuple;
1324                 }
1325         }
1326
1327         /*
1328          * Check the constraints of the tuple
1329          */
1330         if (resultRelationDesc->rd_att->constr)
1331                 ExecConstraints("ExecInsert", resultRelInfo, slot, estate);
1332
1333         /*
1334          * insert the tuple
1335          */
1336         newId = heap_insert(resultRelationDesc, tuple,
1337                                                 estate->es_snapshot->curcid);
1338
1339         IncrAppended();
1340         (estate->es_processed)++;
1341         estate->es_lastoid = newId;
1342         setLastTid(&(tuple->t_self));
1343
1344         /*
1345          * process indices
1346          *
1347          * Note: heap_insert adds a new tuple to a relation.  As a side effect,
1348          * the tupleid of the new tuple is placed in the new tuple's t_ctid
1349          * field.
1350          */
1351         numIndices = resultRelInfo->ri_NumIndices;
1352         if (numIndices > 0)
1353                 ExecInsertIndexTuples(slot, &(tuple->t_self), estate, false);
1354
1355         /* AFTER ROW INSERT Triggers */
1356         ExecARInsertTriggers(estate, resultRelInfo, tuple);
1357 }
1358
1359 /* ----------------------------------------------------------------
1360  *              ExecDelete
1361  *
1362  *              DELETE is like UPDATE, we delete the tuple and its
1363  *              index tuples.
1364  * ----------------------------------------------------------------
1365  */
1366 static void
1367 ExecDelete(TupleTableSlot *slot,
1368                    ItemPointer tupleid,
1369                    EState *estate)
1370 {
1371         ResultRelInfo *resultRelInfo;
1372         Relation        resultRelationDesc;
1373         ItemPointerData ctid;
1374         int                     result;
1375
1376         /*
1377          * get information on the (current) result relation
1378          */
1379         resultRelInfo = estate->es_result_relation_info;
1380         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1381
1382         /* BEFORE ROW DELETE Triggers */
1383         if (resultRelInfo->ri_TrigDesc &&
1384           resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_DELETE] > 0)
1385         {
1386                 bool            dodelete;
1387
1388                 dodelete = ExecBRDeleteTriggers(estate, resultRelInfo, tupleid,
1389                                                                                 estate->es_snapshot->curcid);
1390
1391                 if (!dodelete)                  /* "do nothing" */
1392                         return;
1393         }
1394
1395         /*
1396          * delete the tuple
1397          */
1398 ldelete:;
1399         result = heap_delete(resultRelationDesc, tupleid,
1400                                                  &ctid,
1401                                                  estate->es_snapshot->curcid);
1402         switch (result)
1403         {
1404                 case HeapTupleSelfUpdated:
1405                         /* already deleted by self; nothing to do */
1406                         return;
1407
1408                 case HeapTupleMayBeUpdated:
1409                         break;
1410
1411                 case HeapTupleUpdated:
1412                         if (XactIsoLevel == XACT_SERIALIZABLE)
1413                                 elog(ERROR, "Can't serialize access due to concurrent update");
1414                         else if (!(ItemPointerEquals(tupleid, &ctid)))
1415                         {
1416                                 TupleTableSlot *epqslot = EvalPlanQual(estate,
1417                                                            resultRelInfo->ri_RangeTableIndex, &ctid);
1418
1419                                 if (!TupIsNull(epqslot))
1420                                 {
1421                                         *tupleid = ctid;
1422                                         goto ldelete;
1423                                 }
1424                         }
1425                         /* tuple already deleted; nothing to do */
1426                         return;
1427
1428                 default:
1429                         elog(ERROR, "Unknown status %u from heap_delete", result);
1430                         return;
1431         }
1432
1433         IncrDeleted();
1434         (estate->es_processed)++;
1435
1436         /*
1437          * Note: Normally one would think that we have to delete index tuples
1438          * associated with the heap tuple now..
1439          *
1440          * ... but in POSTGRES, we have no need to do this because the vacuum
1441          * daemon automatically opens an index scan and deletes index tuples
1442          * when it finds deleted heap tuples. -cim 9/27/89
1443          */
1444
1445         /* AFTER ROW DELETE Triggers */
1446         ExecARDeleteTriggers(estate, resultRelInfo, tupleid);
1447 }
1448
1449 /* ----------------------------------------------------------------
1450  *              ExecUpdate
1451  *
1452  *              note: we can't run UPDATE queries with transactions
1453  *              off because UPDATEs are actually INSERTs and our
1454  *              scan will mistakenly loop forever, updating the tuple
1455  *              it just inserted..      This should be fixed but until it
1456  *              is, we don't want to get stuck in an infinite loop
1457  *              which corrupts your database..
1458  * ----------------------------------------------------------------
1459  */
1460 static void
1461 ExecUpdate(TupleTableSlot *slot,
1462                    ItemPointer tupleid,
1463                    EState *estate)
1464 {
1465         HeapTuple       tuple;
1466         ResultRelInfo *resultRelInfo;
1467         Relation        resultRelationDesc;
1468         ItemPointerData ctid;
1469         int                     result;
1470         int                     numIndices;
1471
1472         /*
1473          * abort the operation if not running transactions
1474          */
1475         if (IsBootstrapProcessingMode())
1476         {
1477                 elog(WARNING, "ExecUpdate: UPDATE can't run without transactions");
1478                 return;
1479         }
1480
1481         /*
1482          * get the heap tuple out of the tuple table slot
1483          */
1484         tuple = slot->val;
1485
1486         /*
1487          * get information on the (current) result relation
1488          */
1489         resultRelInfo = estate->es_result_relation_info;
1490         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1491
1492         /* BEFORE ROW UPDATE Triggers */
1493         if (resultRelInfo->ri_TrigDesc &&
1494           resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_UPDATE] > 0)
1495         {
1496                 HeapTuple       newtuple;
1497
1498                 newtuple = ExecBRUpdateTriggers(estate, resultRelInfo,
1499                                                                                 tupleid, tuple,
1500                                                                                 estate->es_snapshot->curcid);
1501
1502                 if (newtuple == NULL)   /* "do nothing" */
1503                         return;
1504
1505                 if (newtuple != tuple)  /* modified by Trigger(s) */
1506                 {
1507                         /*
1508                          * Insert modified tuple into tuple table slot, replacing the
1509                          * original.  We assume that it was allocated in per-tuple
1510                          * memory context, and therefore will go away by itself. The
1511                          * tuple table slot should not try to clear it.
1512                          */
1513                         ExecStoreTuple(newtuple, slot, InvalidBuffer, false);
1514                         tuple = newtuple;
1515                 }
1516         }
1517
1518         /*
1519          * Check the constraints of the tuple
1520          *
1521          * If we generate a new candidate tuple after EvalPlanQual testing, we
1522          * must loop back here and recheck constraints.  (We don't need to
1523          * redo triggers, however.      If there are any BEFORE triggers then
1524          * trigger.c will have done mark4update to lock the correct tuple, so
1525          * there's no need to do them again.)
1526          */
1527 lreplace:;
1528         if (resultRelationDesc->rd_att->constr)
1529                 ExecConstraints("ExecUpdate", resultRelInfo, slot, estate);
1530
1531         /*
1532          * replace the heap tuple
1533          */
1534         result = heap_update(resultRelationDesc, tupleid, tuple,
1535                                                  &ctid,
1536                                                  estate->es_snapshot->curcid);
1537         switch (result)
1538         {
1539                 case HeapTupleSelfUpdated:
1540                         /* already deleted by self; nothing to do */
1541                         return;
1542
1543                 case HeapTupleMayBeUpdated:
1544                         break;
1545
1546                 case HeapTupleUpdated:
1547                         if (XactIsoLevel == XACT_SERIALIZABLE)
1548                                 elog(ERROR, "Can't serialize access due to concurrent update");
1549                         else if (!(ItemPointerEquals(tupleid, &ctid)))
1550                         {
1551                                 TupleTableSlot *epqslot = EvalPlanQual(estate,
1552                                                            resultRelInfo->ri_RangeTableIndex, &ctid);
1553
1554                                 if (!TupIsNull(epqslot))
1555                                 {
1556                                         *tupleid = ctid;
1557                                         tuple = ExecRemoveJunk(estate->es_junkFilter, epqslot);
1558                                         slot = ExecStoreTuple(tuple,
1559                                                                         estate->es_junkFilter->jf_resultSlot,
1560                                                                                   InvalidBuffer, true);
1561                                         goto lreplace;
1562                                 }
1563                         }
1564                         /* tuple already deleted; nothing to do */
1565                         return;
1566
1567                 default:
1568                         elog(ERROR, "Unknown status %u from heap_update", result);
1569                         return;
1570         }
1571
1572         IncrReplaced();
1573         (estate->es_processed)++;
1574
1575         /*
1576          * Note: instead of having to update the old index tuples associated
1577          * with the heap tuple, all we do is form and insert new index tuples.
1578          * This is because UPDATEs are actually DELETEs and INSERTs and index
1579          * tuple deletion is done automagically by the vacuum daemon. All we
1580          * do is insert new index tuples.  -cim 9/27/89
1581          */
1582
1583         /*
1584          * process indices
1585          *
1586          * heap_update updates a tuple in the base relation by invalidating it
1587          * and then inserting a new tuple to the relation.      As a side effect,
1588          * the tupleid of the new tuple is placed in the new tuple's t_ctid
1589          * field.  So we now insert index tuples using the new tupleid stored
1590          * there.
1591          */
1592
1593         numIndices = resultRelInfo->ri_NumIndices;
1594         if (numIndices > 0)
1595                 ExecInsertIndexTuples(slot, &(tuple->t_self), estate, false);
1596
1597         /* AFTER ROW UPDATE Triggers */
1598         ExecARUpdateTriggers(estate, resultRelInfo, tupleid, tuple);
1599 }
1600
1601 static char *
1602 ExecRelCheck(ResultRelInfo *resultRelInfo,
1603                          TupleTableSlot *slot, EState *estate)
1604 {
1605         Relation        rel = resultRelInfo->ri_RelationDesc;
1606         int                     ncheck = rel->rd_att->constr->num_check;
1607         ConstrCheck *check = rel->rd_att->constr->check;
1608         ExprContext *econtext;
1609         MemoryContext oldContext;
1610         List       *qual;
1611         int                     i;
1612
1613         /*
1614          * If first time through for this result relation, build expression
1615          * nodetrees for rel's constraint expressions.  Keep them in the
1616          * per-query memory context so they'll survive throughout the query.
1617          */
1618         if (resultRelInfo->ri_ConstraintExprs == NULL)
1619         {
1620                 oldContext = MemoryContextSwitchTo(estate->es_query_cxt);
1621                 resultRelInfo->ri_ConstraintExprs =
1622                         (List **) palloc(ncheck * sizeof(List *));
1623                 for (i = 0; i < ncheck; i++)
1624                 {
1625                         qual = (List *) stringToNode(check[i].ccbin);
1626                         resultRelInfo->ri_ConstraintExprs[i] = (List *)
1627                                 ExecPrepareExpr((Expr *) qual, estate);
1628                 }
1629                 MemoryContextSwitchTo(oldContext);
1630         }
1631
1632         /*
1633          * We will use the EState's per-tuple context for evaluating
1634          * constraint expressions (creating it if it's not already there).
1635          */
1636         econtext = GetPerTupleExprContext(estate);
1637
1638         /* Arrange for econtext's scan tuple to be the tuple under test */
1639         econtext->ecxt_scantuple = slot;
1640
1641         /* And evaluate the constraints */
1642         for (i = 0; i < ncheck; i++)
1643         {
1644                 qual = resultRelInfo->ri_ConstraintExprs[i];
1645
1646                 /*
1647                  * NOTE: SQL92 specifies that a NULL result from a constraint
1648                  * expression is not to be treated as a failure.  Therefore, tell
1649                  * ExecQual to return TRUE for NULL.
1650                  */
1651                 if (!ExecQual(qual, econtext, true))
1652                         return check[i].ccname;
1653         }
1654
1655         /* NULL result means no error */
1656         return (char *) NULL;
1657 }
1658
1659 void
1660 ExecConstraints(const char *caller, ResultRelInfo *resultRelInfo,
1661                                 TupleTableSlot *slot, EState *estate)
1662 {
1663         Relation        rel = resultRelInfo->ri_RelationDesc;
1664         HeapTuple       tuple = slot->val;
1665         TupleConstr *constr = rel->rd_att->constr;
1666
1667         Assert(constr);
1668
1669         if (constr->has_not_null)
1670         {
1671                 int                     natts = rel->rd_att->natts;
1672                 int                     attrChk;
1673
1674                 for (attrChk = 1; attrChk <= natts; attrChk++)
1675                 {
1676                         if (rel->rd_att->attrs[attrChk - 1]->attnotnull &&
1677                                 heap_attisnull(tuple, attrChk))
1678                                 elog(ERROR, "%s: Fail to add null value in not null attribute %s",
1679                                          caller, NameStr(rel->rd_att->attrs[attrChk - 1]->attname));
1680                 }
1681         }
1682
1683         if (constr->num_check > 0)
1684         {
1685                 char       *failed;
1686
1687                 if ((failed = ExecRelCheck(resultRelInfo, slot, estate)) != NULL)
1688                         elog(ERROR, "%s: rejected due to CHECK constraint \"%s\" on \"%s\"",
1689                                  caller, failed, RelationGetRelationName(rel));
1690         }
1691 }
1692
1693 /*
1694  * Check a modified tuple to see if we want to process its updated version
1695  * under READ COMMITTED rules.
1696  *
1697  * See backend/executor/README for some info about how this works.
1698  */
1699 TupleTableSlot *
1700 EvalPlanQual(EState *estate, Index rti, ItemPointer tid)
1701 {
1702         evalPlanQual *epq;
1703         EState     *epqstate;
1704         Relation        relation;
1705         HeapTupleData tuple;
1706         HeapTuple       copyTuple = NULL;
1707         bool            endNode;
1708
1709         Assert(rti != 0);
1710
1711         /*
1712          * find relation containing target tuple
1713          */
1714         if (estate->es_result_relation_info != NULL &&
1715                 estate->es_result_relation_info->ri_RangeTableIndex == rti)
1716                 relation = estate->es_result_relation_info->ri_RelationDesc;
1717         else
1718         {
1719                 List       *l;
1720
1721                 relation = NULL;
1722                 foreach(l, estate->es_rowMark)
1723                 {
1724                         if (((execRowMark *) lfirst(l))->rti == rti)
1725                         {
1726                                 relation = ((execRowMark *) lfirst(l))->relation;
1727                                 break;
1728                         }
1729                 }
1730                 if (relation == NULL)
1731                         elog(ERROR, "EvalPlanQual: can't find RTE %d", (int) rti);
1732         }
1733
1734         /*
1735          * fetch tid tuple
1736          *
1737          * Loop here to deal with updated or busy tuples
1738          */
1739         tuple.t_self = *tid;
1740         for (;;)
1741         {
1742                 Buffer          buffer;
1743
1744                 if (heap_fetch(relation, SnapshotDirty, &tuple, &buffer, false, NULL))
1745                 {
1746                         TransactionId xwait = SnapshotDirty->xmax;
1747
1748                         if (TransactionIdIsValid(SnapshotDirty->xmin))
1749                                 elog(ERROR, "EvalPlanQual: t_xmin is uncommitted ?!");
1750
1751                         /*
1752                          * If tuple is being updated by other transaction then we have
1753                          * to wait for its commit/abort.
1754                          */
1755                         if (TransactionIdIsValid(xwait))
1756                         {
1757                                 ReleaseBuffer(buffer);
1758                                 XactLockTableWait(xwait);
1759                                 continue;
1760                         }
1761
1762                         /*
1763                          * We got tuple - now copy it for use by recheck query.
1764                          */
1765                         copyTuple = heap_copytuple(&tuple);
1766                         ReleaseBuffer(buffer);
1767                         break;
1768                 }
1769
1770                 /*
1771                  * Oops! Invalid tuple. Have to check is it updated or deleted.
1772                  * Note that it's possible to get invalid SnapshotDirty->tid if
1773                  * tuple updated by this transaction. Have we to check this ?
1774                  */
1775                 if (ItemPointerIsValid(&(SnapshotDirty->tid)) &&
1776                         !(ItemPointerEquals(&(tuple.t_self), &(SnapshotDirty->tid))))
1777                 {
1778                         /* updated, so look at the updated copy */
1779                         tuple.t_self = SnapshotDirty->tid;
1780                         continue;
1781                 }
1782
1783                 /*
1784                  * Deleted or updated by this transaction; forget it.
1785                  */
1786                 return NULL;
1787         }
1788
1789         /*
1790          * For UPDATE/DELETE we have to return tid of actual row we're
1791          * executing PQ for.
1792          */
1793         *tid = tuple.t_self;
1794
1795         /*
1796          * Need to run a recheck subquery.      Find or create a PQ stack entry.
1797          */
1798         epq = estate->es_evalPlanQual;
1799         endNode = true;
1800
1801         if (epq != NULL && epq->rti == 0)
1802         {
1803                 /* Top PQ stack entry is idle, so re-use it */
1804                 Assert(!(estate->es_useEvalPlan) && epq->next == NULL);
1805                 epq->rti = rti;
1806                 endNode = false;
1807         }
1808
1809         /*
1810          * If this is request for another RTE - Ra, - then we have to check
1811          * wasn't PlanQual requested for Ra already and if so then Ra' row was
1812          * updated again and we have to re-start old execution for Ra and
1813          * forget all what we done after Ra was suspended. Cool? -:))
1814          */
1815         if (epq != NULL && epq->rti != rti &&
1816                 epq->estate->es_evTuple[rti - 1] != NULL)
1817         {
1818                 do
1819                 {
1820                         evalPlanQual *oldepq;
1821
1822                         /* stop execution */
1823                         EvalPlanQualStop(epq);
1824                         /* pop previous PlanQual from the stack */
1825                         oldepq = epq->next;
1826                         Assert(oldepq && oldepq->rti != 0);
1827                         /* push current PQ to freePQ stack */
1828                         oldepq->free = epq;
1829                         epq = oldepq;
1830                         estate->es_evalPlanQual = epq;
1831                 } while (epq->rti != rti);
1832         }
1833
1834         /*
1835          * If we are requested for another RTE then we have to suspend
1836          * execution of current PlanQual and start execution for new one.
1837          */
1838         if (epq == NULL || epq->rti != rti)
1839         {
1840                 /* try to reuse plan used previously */
1841                 evalPlanQual *newepq = (epq != NULL) ? epq->free : NULL;
1842
1843                 if (newepq == NULL)             /* first call or freePQ stack is empty */
1844                 {
1845                         newepq = (evalPlanQual *) palloc0(sizeof(evalPlanQual));
1846                         newepq->free = NULL;
1847                         newepq->estate = NULL;
1848                         newepq->planstate = NULL;
1849                 }
1850                 else
1851                 {
1852                         /* recycle previously used PlanQual */
1853                         Assert(newepq->estate == NULL);
1854                         epq->free = NULL;
1855                 }
1856                 /* push current PQ to the stack */
1857                 newepq->next = epq;
1858                 epq = newepq;
1859                 estate->es_evalPlanQual = epq;
1860                 epq->rti = rti;
1861                 endNode = false;
1862         }
1863
1864         Assert(epq->rti == rti);
1865
1866         /*
1867          * Ok - we're requested for the same RTE.  Unfortunately we still have
1868          * to end and restart execution of the plan, because ExecReScan
1869          * wouldn't ensure that upper plan nodes would reset themselves.  We
1870          * could make that work if insertion of the target tuple were
1871          * integrated with the Param mechanism somehow, so that the upper plan
1872          * nodes know that their children's outputs have changed.
1873          *
1874          * Note that the stack of free evalPlanQual nodes is quite useless at
1875          * the moment, since it only saves us from pallocing/releasing the
1876          * evalPlanQual nodes themselves.  But it will be useful once we
1877          * implement ReScan instead of end/restart for re-using PlanQual nodes.
1878          */
1879         if (endNode)
1880         {
1881                 /* stop execution */
1882                 EvalPlanQualStop(epq);
1883         }
1884
1885         /*
1886          * Initialize new recheck query.
1887          *
1888          * Note: if we were re-using PlanQual plans via ExecReScan, we'd need
1889          * to instead copy down changeable state from the top plan (including
1890          * es_result_relation_info, es_junkFilter) and reset locally changeable
1891          * state in the epq (including es_param_exec_vals, es_evTupleNull).
1892          */
1893         EvalPlanQualStart(epq, estate, epq->next);
1894
1895         /*
1896          * free old RTE' tuple, if any, and store target tuple where
1897          * relation's scan node will see it
1898          */
1899         epqstate = epq->estate;
1900         if (epqstate->es_evTuple[rti - 1] != NULL)
1901                 heap_freetuple(epqstate->es_evTuple[rti - 1]);
1902         epqstate->es_evTuple[rti - 1] = copyTuple;
1903
1904         return EvalPlanQualNext(estate);
1905 }
1906
1907 static TupleTableSlot *
1908 EvalPlanQualNext(EState *estate)
1909 {
1910         evalPlanQual *epq = estate->es_evalPlanQual;
1911         MemoryContext oldcontext;
1912         TupleTableSlot *slot;
1913
1914         Assert(epq->rti != 0);
1915
1916 lpqnext:;
1917         oldcontext = MemoryContextSwitchTo(epq->estate->es_query_cxt);
1918         slot = ExecProcNode(epq->planstate);
1919         MemoryContextSwitchTo(oldcontext);
1920
1921         /*
1922          * No more tuples for this PQ. Continue previous one.
1923          */
1924         if (TupIsNull(slot))
1925         {
1926                 evalPlanQual *oldepq;
1927
1928                 /* stop execution */
1929                 EvalPlanQualStop(epq);
1930                 /* pop old PQ from the stack */
1931                 oldepq = epq->next;
1932                 if (oldepq == NULL)
1933                 {
1934                         /* this is the first (oldest) PQ - mark as free */
1935                         epq->rti = 0;
1936                         estate->es_useEvalPlan = false;
1937                         /* and continue Query execution */
1938                         return (NULL);
1939                 }
1940                 Assert(oldepq->rti != 0);
1941                 /* push current PQ to freePQ stack */
1942                 oldepq->free = epq;
1943                 epq = oldepq;
1944                 estate->es_evalPlanQual = epq;
1945                 goto lpqnext;
1946         }
1947
1948         return (slot);
1949 }
1950
1951 static void
1952 EndEvalPlanQual(EState *estate)
1953 {
1954         evalPlanQual *epq = estate->es_evalPlanQual;
1955
1956         if (epq->rti == 0)                      /* plans already shutdowned */
1957         {
1958                 Assert(epq->next == NULL);
1959                 return;
1960         }
1961
1962         for (;;)
1963         {
1964                 evalPlanQual *oldepq;
1965
1966                 /* stop execution */
1967                 EvalPlanQualStop(epq);
1968                 /* pop old PQ from the stack */
1969                 oldepq = epq->next;
1970                 if (oldepq == NULL)
1971                 {
1972                         /* this is the first (oldest) PQ - mark as free */
1973                         epq->rti = 0;
1974                         estate->es_useEvalPlan = false;
1975                         break;
1976                 }
1977                 Assert(oldepq->rti != 0);
1978                 /* push current PQ to freePQ stack */
1979                 oldepq->free = epq;
1980                 epq = oldepq;
1981                 estate->es_evalPlanQual = epq;
1982         }
1983 }
1984
1985 /*
1986  * Start execution of one level of PlanQual.
1987  *
1988  * This is a cut-down version of ExecutorStart(): we copy some state from
1989  * the top-level estate rather than initializing it fresh.
1990  */
1991 static void
1992 EvalPlanQualStart(evalPlanQual *epq, EState *estate, evalPlanQual *priorepq)
1993 {
1994         EState     *epqstate;
1995         int                     rtsize;
1996         MemoryContext oldcontext;
1997
1998         rtsize = length(estate->es_range_table);
1999
2000         epq->estate = epqstate = CreateExecutorState();
2001
2002         oldcontext = MemoryContextSwitchTo(epqstate->es_query_cxt);
2003
2004         /*
2005          * The epqstates share the top query's copy of unchanging state such
2006          * as the snapshot, rangetable, result-rel info, and external Param info.
2007          * They need their own copies of local state, including a tuple table,
2008          * es_param_exec_vals, etc.
2009          */
2010         epqstate->es_direction = ForwardScanDirection;
2011         epqstate->es_snapshot = estate->es_snapshot;
2012         epqstate->es_range_table = estate->es_range_table;
2013         epqstate->es_result_relations = estate->es_result_relations;
2014         epqstate->es_num_result_relations = estate->es_num_result_relations;
2015         epqstate->es_result_relation_info = estate->es_result_relation_info;
2016         epqstate->es_junkFilter = estate->es_junkFilter;
2017         epqstate->es_into_relation_descriptor = estate->es_into_relation_descriptor;
2018         epqstate->es_param_list_info = estate->es_param_list_info;
2019         if (estate->es_topPlan->nParamExec > 0)
2020                 epqstate->es_param_exec_vals = (ParamExecData *)
2021                         palloc0(estate->es_topPlan->nParamExec * sizeof(ParamExecData));
2022         epqstate->es_rowMark = estate->es_rowMark;
2023         epqstate->es_instrument = estate->es_instrument;
2024         epqstate->es_force_oids = estate->es_force_oids;
2025         epqstate->es_topPlan = estate->es_topPlan;
2026         /*
2027          * Each epqstate must have its own es_evTupleNull state, but
2028          * all the stack entries share es_evTuple state.  This allows
2029          * sub-rechecks to inherit the value being examined by an
2030          * outer recheck.
2031          */
2032         epqstate->es_evTupleNull = (bool *) palloc0(rtsize * sizeof(bool));
2033         if (priorepq == NULL)
2034                 /* first PQ stack entry */
2035                 epqstate->es_evTuple = (HeapTuple *)
2036                         palloc0(rtsize * sizeof(HeapTuple));
2037         else
2038                 /* later stack entries share the same storage */
2039                 epqstate->es_evTuple = priorepq->estate->es_evTuple;
2040
2041         epqstate->es_tupleTable =
2042                 ExecCreateTupleTable(estate->es_tupleTable->size);
2043
2044         epq->planstate = ExecInitNode(estate->es_topPlan, epqstate);
2045
2046         MemoryContextSwitchTo(oldcontext);
2047 }
2048
2049 /*
2050  * End execution of one level of PlanQual.
2051  *
2052  * This is a cut-down version of ExecutorEnd(); basically we want to do most
2053  * of the normal cleanup, but *not* close result relations (which we are
2054  * just sharing from the outer query).
2055  */
2056 static void
2057 EvalPlanQualStop(evalPlanQual *epq)
2058 {
2059         EState     *epqstate = epq->estate;
2060         MemoryContext oldcontext;
2061
2062         oldcontext = MemoryContextSwitchTo(epqstate->es_query_cxt);
2063
2064         ExecEndNode(epq->planstate);
2065
2066         ExecDropTupleTable(epqstate->es_tupleTable, true);
2067         epqstate->es_tupleTable = NULL;
2068
2069         if (epqstate->es_evTuple[epq->rti - 1] != NULL)
2070         {
2071                 heap_freetuple(epqstate->es_evTuple[epq->rti - 1]);
2072                 epqstate->es_evTuple[epq->rti - 1] = NULL;
2073         }
2074
2075         MemoryContextSwitchTo(oldcontext);
2076
2077         FreeExecutorState(epqstate);
2078
2079         epq->estate = NULL;
2080         epq->planstate = NULL;
2081 }