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  11  *      The old ExecutorMain() has been replaced by ExecutorStart(),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 1.273 2006/07/03 22:45:38 tgl Exp $
  30  *
  31  *-------------------------------------------------------------------------
  32  */
  33 #include "postgres.h"
  34
  35 #include "access/heapam.h"
  36 #include "access/reloptions.h"
  37 #include "access/xlog.h"
  38 #include "catalog/heap.h"
  39 #include "catalog/namespace.h"
  40 #include "commands/tablecmds.h"
  41 #include "commands/tablespace.h"
  42 #include "commands/trigger.h"
  43 #include "executor/execdebug.h"
  44 #include "executor/execdefs.h"
  45 #include "executor/instrument.h"
  46 #include "miscadmin.h"
  47 #include "optimizer/clauses.h"
  48 #include "optimizer/var.h"
  49 #include "parser/parsetree.h"
  50 #include "parser/parse_clause.h"
  51 #include "storage/smgr.h"
  52 #include "utils/acl.h"
  53 #include "utils/guc.h"
  54 #include "utils/lsyscache.h"
  55 #include "utils/memutils.h"
  56
  57
  58 typedef struct evalPlanQual
  59 {
  60         Index           rti;
  61         EState     *estate;
  62         PlanState  *planstate;
  63         struct evalPlanQual *next;      /* stack of active PlanQual plans */
  64         struct evalPlanQual *free;      /* list of free PlanQual plans */
  65 } evalPlanQual;
  66
  67 /* decls for local routines only used within this module */
  68 static void InitPlan(QueryDesc *queryDesc, int eflags);
  69 static void initResultRelInfo(ResultRelInfo *resultRelInfo,
  70                                   Index resultRelationIndex,
  71                                   List *rangeTable,
  72                                   CmdType operation,
  73                                   bool doInstrument);
  74 static TupleTableSlot *ExecutePlan(EState *estate, PlanState *planstate,
  75                         CmdType operation,
  76                         long numberTuples,
  77                         ScanDirection direction,
  78                         DestReceiver *dest);
  79 static void ExecSelect(TupleTableSlot *slot,
  80                    DestReceiver *dest,
  81                    EState *estate);
  82 static void ExecInsert(TupleTableSlot *slot, ItemPointer tupleid,
  83                    EState *estate);
  84 static void ExecDelete(TupleTableSlot *slot, ItemPointer tupleid,
  85                    EState *estate);
  86 static void ExecUpdate(TupleTableSlot *slot, ItemPointer tupleid,
  87                    EState *estate);
  88 static TupleTableSlot *EvalPlanQualNext(EState *estate);
  89 static void EndEvalPlanQual(EState *estate);
  90 static void ExecCheckRTEPerms(RangeTblEntry *rte);
  91 static void ExecCheckXactReadOnly(Query *parsetree);
  92 static void EvalPlanQualStart(evalPlanQual *epq, EState *estate,
  93                                   evalPlanQual *priorepq);
  94 static void EvalPlanQualStop(evalPlanQual *epq);
  95
  96 /* end of local decls */
  97
  98
  99 /* ----------------------------------------------------------------
 100  *              ExecutorStart
 101  *
 102  *              This routine must be called at the beginning of any execution of any
 103  *              query plan
 104  *
 105  * Takes a QueryDesc previously created by CreateQueryDesc (it's not real
 106  * clear why we bother to separate the two functions, but...).  The tupDesc
 107  * field of the QueryDesc is filled in to describe the tuples that will be
 108  * returned, and the internal fields (estate and planstate) are set up.
 109  *
 110  * eflags contains flag bits as described in executor.h.
 111  *
 112  * NB: the CurrentMemoryContext when this is called will become the parent
 113  * of the per-query context used for this Executor invocation.
 114  * ----------------------------------------------------------------
 115  */
 116 void
 117 ExecutorStart(QueryDesc *queryDesc, int eflags)
 118 {
 119         EState     *estate;
 120         MemoryContext oldcontext;
 121
 122         /* sanity checks: queryDesc must not be started already */
 123         Assert(queryDesc != NULL);
 124         Assert(queryDesc->estate == NULL);
 125
 126         /*
 127          * If the transaction is read-only, we need to check if any writes are
 128          * planned to non-temporary tables.  EXPLAIN is considered read-only.
 129          */
 130         if (XactReadOnly && !(eflags & EXEC_FLAG_EXPLAIN_ONLY))
 131                 ExecCheckXactReadOnly(queryDesc->parsetree);
 132
 133         /*
 134          * Build EState, switch into per-query memory context for startup.
 135          */
 136         estate = CreateExecutorState();
 137         queryDesc->estate = estate;
 138
 139         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 140
 141         /*
 142          * Fill in parameters, if any, from queryDesc
 143          */
 144         estate->es_param_list_info = queryDesc->params;
 145
 146         if (queryDesc->plantree->nParamExec > 0)
 147                 estate->es_param_exec_vals = (ParamExecData *)
 148                         palloc0(queryDesc->plantree->nParamExec * sizeof(ParamExecData));
 149
 150         /*
 151          * Copy other important information into the EState
 152          */
 153         estate->es_snapshot = queryDesc->snapshot;
 154         estate->es_crosscheck_snapshot = queryDesc->crosscheck_snapshot;
 155         estate->es_instrument = queryDesc->doInstrument;
 156
 157         /*
 158          * Initialize the plan state tree
 159          */
 160         InitPlan(queryDesc, eflags);
 161
 162         MemoryContextSwitchTo(oldcontext);
 163 }
 164
 165 /* ----------------------------------------------------------------
 166  *              ExecutorRun
 167  *
 168  *              This is the main routine of the executor module. It accepts
 169  *              the query descriptor from the traffic cop and executes the
 170  *              query plan.
 171  *
 172  *              ExecutorStart must have been called already.
 173  *
 174  *              If direction is NoMovementScanDirection then nothing is done
 175  *              except to start up/shut down the destination.  Otherwise,
 176  *              we retrieve up to 'count' tuples in the specified direction.
 177  *
 178  *              Note: count = 0 is interpreted as no portal limit, i.e., run to
 179  *              completion.
 180  *
 181  * ----------------------------------------------------------------
 182  */
 183 TupleTableSlot *
 184 ExecutorRun(QueryDesc *queryDesc,
 185                         ScanDirection direction, long count)
 186 {
 187         EState     *estate;
 188         CmdType         operation;
 189         DestReceiver *dest;
 190         TupleTableSlot *result;
 191         MemoryContext oldcontext;
 192
 193         /* sanity checks */
 194         Assert(queryDesc != NULL);
 195
 196         estate = queryDesc->estate;
 197
 198         Assert(estate != NULL);
 199
 200         /*
 201          * Switch into per-query memory context
 202          */
 203         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 204
 205         /*
 206          * extract information from the query descriptor and the query feature.
 207          */
 208         operation = queryDesc->operation;
 209         dest = queryDesc->dest;
 210
 211         /*
 212          * startup tuple receiver
 213          */
 214         estate->es_processed = 0;
 215         estate->es_lastoid = InvalidOid;
 216
 217         (*dest->rStartup) (dest, operation, queryDesc->tupDesc);
 218
 219         /*
 220          * run plan
 221          */
 222         if (ScanDirectionIsNoMovement(direction))
 223                 result = NULL;
 224         else
 225                 result = ExecutePlan(estate,
 226                                                          queryDesc->planstate,
 227                                                          operation,
 228                                                          count,
 229                                                          direction,
 230                                                          dest);
 231
 232         /*
 233          * shutdown receiver
 234          */
 235         (*dest->rShutdown) (dest);
 236
 237         MemoryContextSwitchTo(oldcontext);
 238
 239         return result;
 240 }
 241
 242 /* ----------------------------------------------------------------
 243  *              ExecutorEnd
 244  *
 245  *              This routine must be called at the end of execution of any
 246  *              query plan
 247  * ----------------------------------------------------------------
 248  */
 249 void
 250 ExecutorEnd(QueryDesc *queryDesc)
 251 {
 252         EState     *estate;
 253         MemoryContext oldcontext;
 254
 255         /* sanity checks */
 256         Assert(queryDesc != NULL);
 257
 258         estate = queryDesc->estate;
 259
 260         Assert(estate != NULL);
 261
 262         /*
 263          * Switch into per-query memory context to run ExecEndPlan
 264          */
 265         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 266
 267         ExecEndPlan(queryDesc->planstate, estate);
 268
 269         /*
 270          * Must switch out of context before destroying it
 271          */
 272         MemoryContextSwitchTo(oldcontext);
 273
 274         /*
 275          * Release EState and per-query memory context.  This should release
 276          * everything the executor has allocated.
 277          */
 278         FreeExecutorState(estate);
 279
 280         /* Reset queryDesc fields that no longer point to anything */
 281         queryDesc->tupDesc = NULL;
 282         queryDesc->estate = NULL;
 283         queryDesc->planstate = NULL;
 284 }
 285
 286 /* ----------------------------------------------------------------
 287  *              ExecutorRewind
 288  *
 289  *              This routine may be called on an open queryDesc to rewind it
 290  *              to the start.
 291  * ----------------------------------------------------------------
 292  */
 293 void
 294 ExecutorRewind(QueryDesc *queryDesc)
 295 {
 296         EState     *estate;
 297         MemoryContext oldcontext;
 298
 299         /* sanity checks */
 300         Assert(queryDesc != NULL);
 301
 302         estate = queryDesc->estate;
 303
 304         Assert(estate != NULL);
 305
 306         /* It's probably not sensible to rescan updating queries */
 307         Assert(queryDesc->operation == CMD_SELECT);
 308
 309         /*
 310          * Switch into per-query memory context
 311          */
 312         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 313
 314         /*
 315          * rescan plan
 316          */
 317         ExecReScan(queryDesc->planstate, NULL);
 318
 319         MemoryContextSwitchTo(oldcontext);
 320 }
 321
 322
 323 /*
 324  * ExecCheckRTPerms
 325  *              Check access permissions for all relations listed in a range table.
 326  */
 327 void
 328 ExecCheckRTPerms(List *rangeTable)
 329 {
 330         ListCell   *l;
 331
 332         foreach(l, rangeTable)
 333         {
 334                 RangeTblEntry *rte = lfirst(l);
 335
 336                 ExecCheckRTEPerms(rte);
 337         }
 338 }
 339
 340 /*
 341  * ExecCheckRTEPerms
 342  *              Check access permissions for a single RTE.
 343  */
 344 static void
 345 ExecCheckRTEPerms(RangeTblEntry *rte)
 346 {
 347         AclMode         requiredPerms;
 348         Oid                     relOid;
 349         Oid                     userid;
 350
 351         /*
 352          * Only plain-relation RTEs need to be checked here.  Subquery RTEs are
 353          * checked by ExecInitSubqueryScan if the subquery is still a separate
 354          * subquery --- if it's been pulled up into our query level then the RTEs
 355          * are in our rangetable and will be checked here. Function RTEs are
 356          * checked by init_fcache when the function is prepared for execution.
 357          * Join and special RTEs need no checks.
 358          */
 359         if (rte->rtekind != RTE_RELATION)
 360                 return;
 361
 362         /*
 363          * No work if requiredPerms is empty.
 364          */
 365         requiredPerms = rte->requiredPerms;
 366         if (requiredPerms == 0)
 367                 return;
 368
 369         relOid = rte->relid;
 370
 371         /*
 372          * userid to check as: current user unless we have a setuid indication.
 373          *
 374          * Note: GetUserId() is presently fast enough that there's no harm in
 375          * calling it separately for each RTE.  If that stops being true, we could
 376          * call it once in ExecCheckRTPerms and pass the userid down from there.
 377          * But for now, no need for the extra clutter.
 378          */
 379         userid = rte->checkAsUser ? rte->checkAsUser : GetUserId();
 380
 381         /*
 382          * We must have *all* the requiredPerms bits, so use aclmask not aclcheck.
 383          */
 384         if (pg_class_aclmask(relOid, userid, requiredPerms, ACLMASK_ALL)
 385                 != requiredPerms)
 386                 aclcheck_error(ACLCHECK_NO_PRIV, ACL_KIND_CLASS,
 387                                            get_rel_name(relOid));
 388 }
 389
 390 /*
 391  * Check that the query does not imply any writes to non-temp tables.
 392  */
 393 static void
 394 ExecCheckXactReadOnly(Query *parsetree)
 395 {
 396         ListCell   *l;
 397
 398         /*
 399          * CREATE TABLE AS or SELECT INTO?
 400          *
 401          * XXX should we allow this if the destination is temp?
 402          */
 403         if (parsetree->into != NULL)
 404                 goto fail;
 405
 406         /* Fail if write permissions are requested on any non-temp table */
 407         foreach(l, parsetree->rtable)
 408         {
 409                 RangeTblEntry *rte = lfirst(l);
 410
 411                 if (rte->rtekind == RTE_SUBQUERY)
 412                 {
 413                         ExecCheckXactReadOnly(rte->subquery);
 414                         continue;
 415                 }
 416
 417                 if (rte->rtekind != RTE_RELATION)
 418                         continue;
 419
 420                 if ((rte->requiredPerms & (~ACL_SELECT)) == 0)
 421                         continue;
 422
 423                 if (isTempNamespace(get_rel_namespace(rte->relid)))
 424                         continue;
 425
 426                 goto fail;
 427         }
 428
 429         return;
 430
 431 fail:
 432         ereport(ERROR,
 433                         (errcode(ERRCODE_READ_ONLY_SQL_TRANSACTION),
 434                          errmsg("transaction is read-only")));
 435 }
 436
 437
 438 /* ----------------------------------------------------------------
 439  *              InitPlan
 440  *
 441  *              Initializes the query plan: open files, allocate storage
 442  *              and start up the rule manager
 443  * ----------------------------------------------------------------
 444  */
 445 static void
 446 InitPlan(QueryDesc *queryDesc, int eflags)
 447 {
 448         CmdType         operation = queryDesc->operation;
 449         Query      *parseTree = queryDesc->parsetree;
 450         Plan       *plan = queryDesc->plantree;
 451         EState     *estate = queryDesc->estate;
 452         PlanState  *planstate;
 453         List       *rangeTable;
 454         Relation        intoRelationDesc;
 455         bool            do_select_into;
 456         TupleDesc       tupType;
 457         ListCell   *l;
 458
 459         /*
 460          * Do permissions checks.  It's sufficient to examine the query's top
 461          * rangetable here --- subplan RTEs will be checked during
 462          * ExecInitSubPlan().
 463          */
 464         ExecCheckRTPerms(parseTree->rtable);
 465
 466         /*
 467          * get information from query descriptor
 468          */
 469         rangeTable = parseTree->rtable;
 470
 471         /*
 472          * initialize the node's execution state
 473          */
 474         estate->es_range_table = rangeTable;
 475
 476         /*
 477          * if there is a result relation, initialize result relation stuff
 478          */
 479         if (parseTree->resultRelation != 0 && operation != CMD_SELECT)
 480         {
 481                 List       *resultRelations = parseTree->resultRelations;
 482                 int                     numResultRelations;
 483                 ResultRelInfo *resultRelInfos;
 484
 485                 if (resultRelations != NIL)
 486                 {
 487                         /*
 488                          * Multiple result relations (due to inheritance)
 489                          * parseTree->resultRelations identifies them all
 490                          */
 491                         ResultRelInfo *resultRelInfo;
 492
 493                         numResultRelations = list_length(resultRelations);
 494                         resultRelInfos = (ResultRelInfo *)
 495                                 palloc(numResultRelations * sizeof(ResultRelInfo));
 496                         resultRelInfo = resultRelInfos;
 497                         foreach(l, resultRelations)
 498                         {
 499                                 initResultRelInfo(resultRelInfo,
 500                                                                   lfirst_int(l),
 501                                                                   rangeTable,
 502                                                                   operation,
 503                                                                   estate->es_instrument);
 504                                 resultRelInfo++;
 505                         }
 506                 }
 507                 else
 508                 {
 509                         /*
 510                          * Single result relation identified by parseTree->resultRelation
 511                          */
 512                         numResultRelations = 1;
 513                         resultRelInfos = (ResultRelInfo *) palloc(sizeof(ResultRelInfo));
 514                         initResultRelInfo(resultRelInfos,
 515                                                           parseTree->resultRelation,
 516                                                           rangeTable,
 517                                                           operation,
 518                                                           estate->es_instrument);
 519                 }
 520
 521                 estate->es_result_relations = resultRelInfos;
 522                 estate->es_num_result_relations = numResultRelations;
 523                 /* Initialize to first or only result rel */
 524                 estate->es_result_relation_info = resultRelInfos;
 525         }
 526         else
 527         {
 528                 /*
 529                  * if no result relation, then set state appropriately
 530                  */
 531                 estate->es_result_relations = NULL;
 532                 estate->es_num_result_relations = 0;
 533                 estate->es_result_relation_info = NULL;
 534         }
 535
 536         /*
 537          * Detect whether we're doing SELECT INTO.  If so, set the es_into_oids
 538          * flag appropriately so that the plan tree will be initialized with the
 539          * correct tuple descriptors.
 540          */
 541         do_select_into = false;
 542
 543         if (operation == CMD_SELECT && parseTree->into != NULL)
 544         {
 545                 do_select_into = true;
 546                 estate->es_select_into = true;
 547                 estate->es_into_oids = interpretOidsOption(parseTree->intoOptions);
 548         }
 549
 550         /*
 551          * Have to lock relations selected FOR UPDATE/FOR SHARE
 552          */
 553         estate->es_rowMarks = NIL;
 554         foreach(l, parseTree->rowMarks)
 555         {
 556                 RowMarkClause *rc = (RowMarkClause *) lfirst(l);
 557                 Oid                     relid = getrelid(rc->rti, rangeTable);
 558                 Relation        relation;
 559                 ExecRowMark *erm;
 560
 561                 relation = heap_open(relid, RowShareLock);
 562                 erm = (ExecRowMark *) palloc(sizeof(ExecRowMark));
 563                 erm->relation = relation;
 564                 erm->rti = rc->rti;
 565                 erm->forUpdate = rc->forUpdate;
 566                 erm->noWait = rc->noWait;
 567                 snprintf(erm->resname, sizeof(erm->resname), "ctid%u", rc->rti);
 568                 estate->es_rowMarks = lappend(estate->es_rowMarks, erm);
 569         }
 570
 571         /*
 572          * initialize the executor "tuple" table.  We need slots for all the plan
 573          * nodes, plus possibly output slots for the junkfilter(s). At this point
 574          * we aren't sure if we need junkfilters, so just add slots for them
 575          * unconditionally.  Also, if it's not a SELECT, set up a slot for use for
 576          * trigger output tuples.
 577          */
 578         {
 579                 int                     nSlots = ExecCountSlotsNode(plan);
 580
 581                 if (parseTree->resultRelations != NIL)
 582                         nSlots += list_length(parseTree->resultRelations);
 583                 else
 584                         nSlots += 1;
 585                 if (operation != CMD_SELECT)
 586                         nSlots++;
 587
 588                 estate->es_tupleTable = ExecCreateTupleTable(nSlots);
 589
 590                 if (operation != CMD_SELECT)
 591                         estate->es_trig_tuple_slot =
 592                                 ExecAllocTableSlot(estate->es_tupleTable);
 593         }
 594
 595         /* mark EvalPlanQual not active */
 596         estate->es_topPlan = plan;
 597         estate->es_evalPlanQual = NULL;
 598         estate->es_evTupleNull = NULL;
 599         estate->es_evTuple = NULL;
 600         estate->es_useEvalPlan = false;
 601
 602         /*
 603          * initialize the private state information for all the nodes in the query
 604          * tree.  This opens files, allocates storage and leaves us ready to start
 605          * processing tuples.
 606          */
 607         planstate = ExecInitNode(plan, estate, eflags);
 608
 609         /*
 610          * Get the tuple descriptor describing the type of tuples to return. (this
 611          * is especially important if we are creating a relation with "SELECT
 612          * INTO")
 613          */
 614         tupType = ExecGetResultType(planstate);
 615
 616         /*
 617          * Initialize the junk filter if needed.  SELECT and INSERT queries need a
 618          * filter if there are any junk attrs in the tlist.  INSERT and SELECT
 619          * INTO also need a filter if the plan may return raw disk tuples (else
 620          * heap_insert will be scribbling on the source relation!). UPDATE and
 621          * DELETE always need a filter, since there's always a junk 'ctid'
 622          * attribute present --- no need to look first.
 623          */
 624         {
 625                 bool            junk_filter_needed = false;
 626                 ListCell   *tlist;
 627
 628                 switch (operation)
 629                 {
 630                         case CMD_SELECT:
 631                         case CMD_INSERT:
 632                                 foreach(tlist, plan->targetlist)
 633                                 {
 634                                         TargetEntry *tle = (TargetEntry *) lfirst(tlist);
 635
 636                                         if (tle->resjunk)
 637                                         {
 638                                                 junk_filter_needed = true;
 639                                                 break;
 640                                         }
 641                                 }
 642                                 if (!junk_filter_needed &&
 643                                         (operation == CMD_INSERT || do_select_into) &&
 644                                         ExecMayReturnRawTuples(planstate))
 645                                         junk_filter_needed = true;
 646                                 break;
 647                         case CMD_UPDATE:
 648                         case CMD_DELETE:
 649                                 junk_filter_needed = true;
 650                                 break;
 651                         default:
 652                                 break;
 653                 }
 654
 655                 if (junk_filter_needed)
 656                 {
 657                         /*
 658                          * If there are multiple result relations, each one needs its own
 659                          * junk filter.  Note this is only possible for UPDATE/DELETE, so
 660                          * we can't be fooled by some needing a filter and some not.
 661                          */
 662                         if (parseTree->resultRelations != NIL)
 663                         {
 664                                 PlanState **appendplans;
 665                                 int                     as_nplans;
 666                                 ResultRelInfo *resultRelInfo;
 667                                 int                     i;
 668
 669                                 /* Top plan had better be an Append here. */
 670                                 Assert(IsA(plan, Append));
 671                                 Assert(((Append *) plan)->isTarget);
 672                                 Assert(IsA(planstate, AppendState));
 673                                 appendplans = ((AppendState *) planstate)->appendplans;
 674                                 as_nplans = ((AppendState *) planstate)->as_nplans;
 675                                 Assert(as_nplans == estate->es_num_result_relations);
 676                                 resultRelInfo = estate->es_result_relations;
 677                                 for (i = 0; i < as_nplans; i++)
 678                                 {
 679                                         PlanState  *subplan = appendplans[i];
 680                                         JunkFilter *j;
 681
 682                                         j = ExecInitJunkFilter(subplan->plan->targetlist,
 683                                                         resultRelInfo->ri_RelationDesc->rd_att->tdhasoid,
 684                                                                   ExecAllocTableSlot(estate->es_tupleTable));
 685                                         resultRelInfo->ri_junkFilter = j;
 686                                         resultRelInfo++;
 687                                 }
 688
 689                                 /*
 690                                  * Set active junkfilter too; at this point ExecInitAppend has
 691                                  * already selected an active result relation...
 692                                  */
 693                                 estate->es_junkFilter =
 694                                         estate->es_result_relation_info->ri_junkFilter;
 695                         }
 696                         else
 697                         {
 698                                 /* Normal case with just one JunkFilter */
 699                                 JunkFilter *j;
 700
 701                                 j = ExecInitJunkFilter(planstate->plan->targetlist,
 702                                                                            tupType->tdhasoid,
 703                                                                   ExecAllocTableSlot(estate->es_tupleTable));
 704                                 estate->es_junkFilter = j;
 705                                 if (estate->es_result_relation_info)
 706                                         estate->es_result_relation_info->ri_junkFilter = j;
 707
 708                                 /* For SELECT, want to return the cleaned tuple type */
 709                                 if (operation == CMD_SELECT)
 710                                         tupType = j->jf_cleanTupType;
 711                         }
 712                 }
 713                 else
 714                         estate->es_junkFilter = NULL;
 715         }
 716
 717         /*
 718          * If doing SELECT INTO, initialize the "into" relation.  We must wait
 719          * till now so we have the "clean" result tuple type to create the new
 720          * table from.
 721          *
 722          * If EXPLAIN, skip creating the "into" relation.
 723          */
 724         intoRelationDesc = NULL;
 725
 726         if (do_select_into && !(eflags & EXEC_FLAG_EXPLAIN_ONLY))
 727         {
 728                 char       *intoName;
 729                 Oid                     namespaceId;
 730                 Oid                     tablespaceId;
 731                 Datum           reloptions;
 732                 AclResult       aclresult;
 733                 Oid                     intoRelationId;
 734                 TupleDesc       tupdesc;
 735
 736                 /*
 737                  * Check consistency of arguments
 738                  */
 739                 if (parseTree->intoOnCommit != ONCOMMIT_NOOP && !parseTree->into->istemp)
 740                         ereport(ERROR,
 741                                         (errcode(ERRCODE_INVALID_TABLE_DEFINITION),
 742                                          errmsg("ON COMMIT can only be used on temporary tables")));
 743
 744                 /*
 745                  * find namespace to create in, check permissions
 746                  */
 747                 intoName = parseTree->into->relname;
 748                 namespaceId = RangeVarGetCreationNamespace(parseTree->into);
 749
 750                 aclresult = pg_namespace_aclcheck(namespaceId, GetUserId(),
 751                                                                                   ACL_CREATE);
 752                 if (aclresult != ACLCHECK_OK)
 753                         aclcheck_error(aclresult, ACL_KIND_NAMESPACE,
 754                                                    get_namespace_name(namespaceId));
 755
 756                 /*
 757                  * Select tablespace to use.  If not specified, use default_tablespace
 758                  * (which may in turn default to database's default).
 759                  */
 760                 if (parseTree->intoTableSpaceName)
 761                 {
 762                         tablespaceId = get_tablespace_oid(parseTree->intoTableSpaceName);
 763                         if (!OidIsValid(tablespaceId))
 764                                 ereport(ERROR,
 765                                                 (errcode(ERRCODE_UNDEFINED_OBJECT),
 766                                                  errmsg("tablespace \"%s\" does not exist",
 767                                                                 parseTree->intoTableSpaceName)));
 768                 } else
 769                 {
 770                         tablespaceId = GetDefaultTablespace();
 771                         /* note InvalidOid is OK in this case */
 772                 }
 773
 774                 /* Parse and validate any reloptions */
 775                 reloptions = transformRelOptions((Datum) 0,
 776                                                                                  parseTree->intoOptions,
 777                                                                                  true,
 778                                                                                  false);
 779                 (void) heap_reloptions(RELKIND_RELATION, reloptions, true);
 780
 781                 /* Check permissions except when using the database's default */
 782                 if (OidIsValid(tablespaceId))
 783                 {
 784                         AclResult       aclresult;
 785
 786                         aclresult = pg_tablespace_aclcheck(tablespaceId, GetUserId(),
 787                                                                                            ACL_CREATE);
 788
 789                         if (aclresult != ACLCHECK_OK)
 790                                 aclcheck_error(aclresult, ACL_KIND_TABLESPACE,
 791                                                            get_tablespace_name(tablespaceId));
 792                 }
 793
 794                 /*
 795                  * have to copy tupType to get rid of constraints
 796                  */
 797                 tupdesc = CreateTupleDescCopy(tupType);
 798
 799                 intoRelationId = heap_create_with_catalog(intoName,
 800                                                                                                   namespaceId,
 801                                                                                                   tablespaceId,
 802                                                                                                   InvalidOid,
 803                                                                                                   GetUserId(),
 804                                                                                                   tupdesc,
 805                                                                                                   RELKIND_RELATION,
 806                                                                                                   false,
 807                                                                                                   true,
 808                                                                                                   0,
 809                                                                                                   parseTree->intoOnCommit,
 810                                                                                                   reloptions,
 811                                                                                                   allowSystemTableMods);
 812
 813                 FreeTupleDesc(tupdesc);
 814
 815                 /*
 816                  * Advance command counter so that the newly-created relation's
 817                  * catalog tuples will be visible to heap_open.
 818                  */
 819                 CommandCounterIncrement();
 820
 821                 /*
 822                  * If necessary, create a TOAST table for the into relation. Note that
 823                  * AlterTableCreateToastTable ends with CommandCounterIncrement(), so
 824                  * that the TOAST table will be visible for insertion.
 825                  */
 826                 AlterTableCreateToastTable(intoRelationId, true);
 827
 828                 /*
 829                  * And open the constructed table for writing.
 830                  */
 831                 intoRelationDesc = heap_open(intoRelationId, AccessExclusiveLock);
 832
 833                 /* use_wal off requires rd_targblock be initially invalid */
 834                 Assert(intoRelationDesc->rd_targblock == InvalidBlockNumber);
 835
 836                 /*
 837                  * We can skip WAL-logging the insertions, unless PITR is in use.
 838                  *
 839                  * Note that for a non-temp INTO table, this is safe only because we
 840                  * know that the catalog changes above will have been WAL-logged, and
 841                  * so RecordTransactionCommit will think it needs to WAL-log the
 842                  * eventual transaction commit.  Else the commit might be lost, even
 843                  * though all the data is safely fsync'd ...
 844                  */
 845                 estate->es_into_relation_use_wal = XLogArchivingActive();
 846         }
 847
 848         estate->es_into_relation_descriptor = intoRelationDesc;
 849
 850         queryDesc->tupDesc = tupType;
 851         queryDesc->planstate = planstate;
 852 }
 853
 854 /*
 855  * Initialize ResultRelInfo data for one result relation
 856  */
 857 static void
 858 initResultRelInfo(ResultRelInfo *resultRelInfo,
 859                                   Index resultRelationIndex,
 860                                   List *rangeTable,
 861                                   CmdType operation,
 862                                   bool doInstrument)
 863 {
 864         Oid                     resultRelationOid;
 865         Relation        resultRelationDesc;
 866
 867         resultRelationOid = getrelid(resultRelationIndex, rangeTable);
 868         resultRelationDesc = heap_open(resultRelationOid, RowExclusiveLock);
 869
 870         switch (resultRelationDesc->rd_rel->relkind)
 871         {
 872                 case RELKIND_SEQUENCE:
 873                         ereport(ERROR,
 874                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 875                                          errmsg("cannot change sequence \"%s\"",
 876                                                         RelationGetRelationName(resultRelationDesc))));
 877                         break;
 878                 case RELKIND_TOASTVALUE:
 879                         ereport(ERROR,
 880                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 881                                          errmsg("cannot change TOAST relation \"%s\"",
 882                                                         RelationGetRelationName(resultRelationDesc))));
 883                         break;
 884                 case RELKIND_VIEW:
 885                         ereport(ERROR,
 886                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 887                                          errmsg("cannot change view \"%s\"",
 888                                                         RelationGetRelationName(resultRelationDesc))));
 889                         break;
 890         }
 891
 892         MemSet(resultRelInfo, 0, sizeof(ResultRelInfo));
 893         resultRelInfo->type = T_ResultRelInfo;
 894         resultRelInfo->ri_RangeTableIndex = resultRelationIndex;
 895         resultRelInfo->ri_RelationDesc = resultRelationDesc;
 896         resultRelInfo->ri_NumIndices = 0;
 897         resultRelInfo->ri_IndexRelationDescs = NULL;
 898         resultRelInfo->ri_IndexRelationInfo = NULL;
 899         /* make a copy so as not to depend on relcache info not changing... */
 900         resultRelInfo->ri_TrigDesc = CopyTriggerDesc(resultRelationDesc->trigdesc);
 901         if (resultRelInfo->ri_TrigDesc)
 902         {
 903                 int                     n = resultRelInfo->ri_TrigDesc->numtriggers;
 904
 905                 resultRelInfo->ri_TrigFunctions = (FmgrInfo *)
 906                         palloc0(n * sizeof(FmgrInfo));
 907                 if (doInstrument)
 908                         resultRelInfo->ri_TrigInstrument = InstrAlloc(n);
 909                 else
 910                         resultRelInfo->ri_TrigInstrument = NULL;
 911         }
 912         else
 913         {
 914                 resultRelInfo->ri_TrigFunctions = NULL;
 915                 resultRelInfo->ri_TrigInstrument = NULL;
 916         }
 917         resultRelInfo->ri_ConstraintExprs = NULL;
 918         resultRelInfo->ri_junkFilter = NULL;
 919
 920         /*
 921          * If there are indices on the result relation, open them and save
 922          * descriptors in the result relation info, so that we can add new index
 923          * entries for the tuples we add/update.  We need not do this for a
 924          * DELETE, however, since deletion doesn't affect indexes.
 925          */
 926         if (resultRelationDesc->rd_rel->relhasindex &&
 927                 operation != CMD_DELETE)
 928                 ExecOpenIndices(resultRelInfo);
 929 }
 930
 931 /*
 932  *              ExecContextForcesOids
 933  *
 934  * This is pretty grotty: when doing INSERT, UPDATE, or SELECT INTO,
 935  * we need to ensure that result tuples have space for an OID iff they are
 936  * going to be stored into a relation that has OIDs.  In other contexts
 937  * we are free to choose whether to leave space for OIDs in result tuples
 938  * (we generally don't want to, but we do if a physical-tlist optimization
 939  * is possible).  This routine checks the plan context and returns TRUE if the
 940  * choice is forced, FALSE if the choice is not forced.  In the TRUE case,
 941  * *hasoids is set to the required value.
 942  *
 943  * One reason this is ugly is that all plan nodes in the plan tree will emit
 944  * tuples with space for an OID, though we really only need the topmost node
 945  * to do so.  However, node types like Sort don't project new tuples but just
 946  * return their inputs, and in those cases the requirement propagates down
 947  * to the input node.  Eventually we might make this code smart enough to
 948  * recognize how far down the requirement really goes, but for now we just
 949  * make all plan nodes do the same thing if the top level forces the choice.
 950  *
 951  * We assume that estate->es_result_relation_info is already set up to
 952  * describe the target relation.  Note that in an UPDATE that spans an
 953  * inheritance tree, some of the target relations may have OIDs and some not.
 954  * We have to make the decisions on a per-relation basis as we initialize
 955  * each of the child plans of the topmost Append plan.
 956  *
 957  * SELECT INTO is even uglier, because we don't have the INTO relation's
 958  * descriptor available when this code runs; we have to look aside at a
 959  * flag set by InitPlan().
 960  */
 961 bool
 962 ExecContextForcesOids(PlanState *planstate, bool *hasoids)
 963 {
 964         if (planstate->state->es_select_into)
 965         {
 966                 *hasoids = planstate->state->es_into_oids;
 967                 return true;
 968         }
 969         else
 970         {
 971                 ResultRelInfo *ri = planstate->state->es_result_relation_info;
 972
 973                 if (ri != NULL)
 974                 {
 975                         Relation        rel = ri->ri_RelationDesc;
 976
 977                         if (rel != NULL)
 978                         {
 979                                 *hasoids = rel->rd_rel->relhasoids;
 980                                 return true;
 981                         }
 982                 }
 983         }
 984
 985         return false;
 986 }
 987
 988 /* ----------------------------------------------------------------
 989  *              ExecEndPlan
 990  *
 991  *              Cleans up the query plan -- closes files and frees up storage
 992  *
 993  * NOTE: we are no longer very worried about freeing storage per se
 994  * in this code; FreeExecutorState should be guaranteed to release all
 995  * memory that needs to be released.  What we are worried about doing
 996  * is closing relations and dropping buffer pins.  Thus, for example,
 997  * tuple tables must be cleared or dropped to ensure pins are released.
 998  * ----------------------------------------------------------------
 999  */
1000 void
1001 ExecEndPlan(PlanState *planstate, EState *estate)
1002 {
1003         ResultRelInfo *resultRelInfo;
1004         int                     i;
1005         ListCell   *l;
1006
1007         /*
1008          * shut down any PlanQual processing we were doing
1009          */
1010         if (estate->es_evalPlanQual != NULL)
1011                 EndEvalPlanQual(estate);
1012
1013         /*
1014          * shut down the node-type-specific query processing
1015          */
1016         ExecEndNode(planstate);
1017
1018         /*
1019          * destroy the executor "tuple" table.
1020          */
1021         ExecDropTupleTable(estate->es_tupleTable, true);
1022         estate->es_tupleTable = NULL;
1023
1024         /*
1025          * close the result relation(s) if any, but hold locks until xact commit.
1026          */
1027         resultRelInfo = estate->es_result_relations;
1028         for (i = estate->es_num_result_relations; i > 0; i--)
1029         {
1030                 /* Close indices and then the relation itself */
1031                 ExecCloseIndices(resultRelInfo);
1032                 heap_close(resultRelInfo->ri_RelationDesc, NoLock);
1033                 resultRelInfo++;
1034         }
1035
1036         /*
1037          * close the "into" relation if necessary, again keeping lock
1038          */
1039         if (estate->es_into_relation_descriptor != NULL)
1040         {
1041                 /*
1042                  * If we skipped using WAL, and it's not a temp relation, we must
1043                  * force the relation down to disk before it's safe to commit the
1044                  * transaction.  This requires forcing out any dirty buffers and then
1045                  * doing a forced fsync.
1046                  */
1047                 if (!estate->es_into_relation_use_wal &&
1048                         !estate->es_into_relation_descriptor->rd_istemp)
1049                 {
1050                         FlushRelationBuffers(estate->es_into_relation_descriptor);
1051                         /* FlushRelationBuffers will have opened rd_smgr */
1052                         smgrimmedsync(estate->es_into_relation_descriptor->rd_smgr);
1053                 }
1054
1055                 heap_close(estate->es_into_relation_descriptor, NoLock);
1056         }
1057
1058         /*
1059          * close any relations selected FOR UPDATE/FOR SHARE, again keeping locks
1060          */
1061         foreach(l, estate->es_rowMarks)
1062         {
1063                 ExecRowMark *erm = lfirst(l);
1064
1065                 heap_close(erm->relation, NoLock);
1066         }
1067 }
1068
1069 /* ----------------------------------------------------------------
1070  *              ExecutePlan
1071  *
1072  *              processes the query plan to retrieve 'numberTuples' tuples in the
1073  *              direction specified.
1074  *
1075  *              Retrieves all tuples if numberTuples is 0
1076  *
1077  *              result is either a slot containing the last tuple in the case
1078  *              of a SELECT or NULL otherwise.
1079  *
1080  * Note: the ctid attribute is a 'junk' attribute that is removed before the
1081  * user can see it
1082  * ----------------------------------------------------------------
1083  */
1084 static TupleTableSlot *
1085 ExecutePlan(EState *estate,
1086                         PlanState *planstate,
1087                         CmdType operation,
1088                         long numberTuples,
1089                         ScanDirection direction,
1090                         DestReceiver *dest)
1091 {
1092         JunkFilter *junkfilter;
1093         TupleTableSlot *slot;
1094         ItemPointer tupleid = NULL;
1095         ItemPointerData tuple_ctid;
1096         long            current_tuple_count;
1097         TupleTableSlot *result;
1098
1099         /*
1100          * initialize local variables
1101          */
1102         slot = NULL;
1103         current_tuple_count = 0;
1104         result = NULL;
1105
1106         /*
1107          * Set the direction.
1108          */
1109         estate->es_direction = direction;
1110
1111         /*
1112          * Process BEFORE EACH STATEMENT triggers
1113          */
1114         switch (operation)
1115         {
1116                 case CMD_UPDATE:
1117                         ExecBSUpdateTriggers(estate, estate->es_result_relation_info);
1118                         break;
1119                 case CMD_DELETE:
1120                         ExecBSDeleteTriggers(estate, estate->es_result_relation_info);
1121                         break;
1122                 case CMD_INSERT:
1123                         ExecBSInsertTriggers(estate, estate->es_result_relation_info);
1124                         break;
1125                 default:
1126                         /* do nothing */
1127                         break;
1128         }
1129
1130         /*
1131          * Loop until we've processed the proper number of tuples from the plan.
1132          */
1133
1134         for (;;)
1135         {
1136                 /* Reset the per-output-tuple exprcontext */
1137                 ResetPerTupleExprContext(estate);
1138
1139                 /*
1140                  * Execute the plan and obtain a tuple
1141                  */
1142 lnext:  ;
1143                 if (estate->es_useEvalPlan)
1144                 {
1145                         slot = EvalPlanQualNext(estate);
1146                         if (TupIsNull(slot))
1147                                 slot = ExecProcNode(planstate);
1148                 }
1149                 else
1150                         slot = ExecProcNode(planstate);
1151
1152                 /*
1153                  * if the tuple is null, then we assume there is nothing more to
1154                  * process so we just return null...
1155                  */
1156                 if (TupIsNull(slot))
1157                 {
1158                         result = NULL;
1159                         break;
1160                 }
1161
1162                 /*
1163                  * if we have a junk filter, then project a new tuple with the junk
1164                  * removed.
1165                  *
1166                  * Store this new "clean" tuple in the junkfilter's resultSlot.
1167                  * (Formerly, we stored it back over the "dirty" tuple, which is WRONG
1168                  * because that tuple slot has the wrong descriptor.)
1169                  *
1170                  * Also, extract all the junk information we need.
1171                  */
1172                 if ((junkfilter = estate->es_junkFilter) != NULL)
1173                 {
1174                         Datum           datum;
1175                         bool            isNull;
1176
1177                         /*
1178                          * extract the 'ctid' junk attribute.
1179                          */
1180                         if (operation == CMD_UPDATE || operation == CMD_DELETE)
1181                         {
1182                                 if (!ExecGetJunkAttribute(junkfilter,
1183                                                                                   slot,
1184                                                                                   "ctid",
1185                                                                                   &datum,
1186                                                                                   &isNull))
1187                                         elog(ERROR, "could not find junk ctid column");
1188
1189                                 /* shouldn't ever get a null result... */
1190                                 if (isNull)
1191                                         elog(ERROR, "ctid is NULL");
1192
1193                                 tupleid = (ItemPointer) DatumGetPointer(datum);
1194                                 tuple_ctid = *tupleid;  /* make sure we don't free the ctid!! */
1195                                 tupleid = &tuple_ctid;
1196                         }
1197
1198                         /*
1199                          * Process any FOR UPDATE or FOR SHARE locking requested.
1200                          */
1201                         else if (estate->es_rowMarks != NIL)
1202                         {
1203                                 ListCell   *l;
1204
1205                 lmark:  ;
1206                                 foreach(l, estate->es_rowMarks)
1207                                 {
1208                                         ExecRowMark *erm = lfirst(l);
1209                                         HeapTupleData tuple;
1210                                         Buffer          buffer;
1211                                         ItemPointerData update_ctid;
1212                                         TransactionId update_xmax;
1213                                         TupleTableSlot *newSlot;
1214                                         LockTupleMode lockmode;
1215                                         HTSU_Result test;
1216
1217                                         if (!ExecGetJunkAttribute(junkfilter,
1218                                                                                           slot,
1219                                                                                           erm->resname,
1220                                                                                           &datum,
1221                                                                                           &isNull))
1222                                                 elog(ERROR, "could not find junk \"%s\" column",
1223                                                          erm->resname);
1224
1225                                         /* shouldn't ever get a null result... */
1226                                         if (isNull)
1227                                                 elog(ERROR, "\"%s\" is NULL", erm->resname);
1228
1229                                         tuple.t_self = *((ItemPointer) DatumGetPointer(datum));
1230
1231                                         if (erm->forUpdate)
1232                                                 lockmode = LockTupleExclusive;
1233                                         else
1234                                                 lockmode = LockTupleShared;
1235
1236                                         test = heap_lock_tuple(erm->relation, &tuple, &buffer,
1237                                                                                    &update_ctid, &update_xmax,
1238                                                                                    estate->es_snapshot->curcid,
1239                                                                                    lockmode, erm->noWait);
1240                                         ReleaseBuffer(buffer);
1241                                         switch (test)
1242                                         {
1243                                                 case HeapTupleSelfUpdated:
1244                                                         /* treat it as deleted; do not process */
1245                                                         goto lnext;
1246
1247                                                 case HeapTupleMayBeUpdated:
1248                                                         break;
1249
1250                                                 case HeapTupleUpdated:
1251                                                         if (IsXactIsoLevelSerializable)
1252                                                                 ereport(ERROR,
1253                                                                  (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1254                                                                   errmsg("could not serialize access due to concurrent update")));
1255                                                         if (!ItemPointerEquals(&update_ctid,
1256                                                                                                    &tuple.t_self))
1257                                                         {
1258                                                                 /* updated, so look at updated version */
1259                                                                 newSlot = EvalPlanQual(estate,
1260                                                                                                            erm->rti,
1261                                                                                                            &update_ctid,
1262                                                                                                            update_xmax,
1263                                                                                                            estate->es_snapshot->curcid);
1264                                                                 if (!TupIsNull(newSlot))
1265                                                                 {
1266                                                                         slot = newSlot;
1267                                                                         estate->es_useEvalPlan = true;
1268                                                                         goto lmark;
1269                                                                 }
1270                                                         }
1271
1272                                                         /*
1273                                                          * if tuple was deleted or PlanQual failed for
1274                                                          * updated tuple - we must not return this tuple!
1275                                                          */
1276                                                         goto lnext;
1277
1278                                                 default:
1279                                                         elog(ERROR, "unrecognized heap_lock_tuple status: %u",
1280                                                                  test);
1281                                                         return NULL;
1282                                         }
1283                                 }
1284                         }
1285
1286                         /*
1287                          * Finally create a new "clean" tuple with all junk attributes
1288                          * removed
1289                          */
1290                         slot = ExecFilterJunk(junkfilter, slot);
1291                 }
1292
1293                 /*
1294                  * now that we have a tuple, do the appropriate thing with it.. either
1295                  * return it to the user, add it to a relation someplace, delete it
1296                  * from a relation, or modify some of its attributes.
1297                  */
1298                 switch (operation)
1299                 {
1300                         case CMD_SELECT:
1301                                 ExecSelect(slot,        /* slot containing tuple */
1302                                                    dest,        /* destination's tuple-receiver obj */
1303                                                    estate);
1304                                 result = slot;
1305                                 break;
1306
1307                         case CMD_INSERT:
1308                                 ExecInsert(slot, tupleid, estate);
1309                                 result = NULL;
1310                                 break;
1311
1312                         case CMD_DELETE:
1313                                 ExecDelete(slot, tupleid, estate);
1314                                 result = NULL;
1315                                 break;
1316
1317                         case CMD_UPDATE:
1318                                 ExecUpdate(slot, tupleid, estate);
1319                                 result = NULL;
1320                                 break;
1321
1322                         default:
1323                                 elog(ERROR, "unrecognized operation code: %d",
1324                                          (int) operation);
1325                                 result = NULL;
1326                                 break;
1327                 }
1328
1329                 /*
1330                  * check our tuple count.. if we've processed the proper number then
1331                  * quit, else loop again and process more tuples.  Zero numberTuples
1332                  * means no limit.
1333                  */
1334                 current_tuple_count++;
1335                 if (numberTuples && numberTuples == current_tuple_count)
1336                         break;
1337         }
1338
1339         /*
1340          * Process AFTER EACH STATEMENT triggers
1341          */
1342         switch (operation)
1343         {
1344                 case CMD_UPDATE:
1345                         ExecASUpdateTriggers(estate, estate->es_result_relation_info);
1346                         break;
1347                 case CMD_DELETE:
1348                         ExecASDeleteTriggers(estate, estate->es_result_relation_info);
1349                         break;
1350                 case CMD_INSERT:
1351                         ExecASInsertTriggers(estate, estate->es_result_relation_info);
1352                         break;
1353                 default:
1354                         /* do nothing */
1355                         break;
1356         }
1357
1358         /*
1359          * here, result is either a slot containing a tuple in the case of a
1360          * SELECT or NULL otherwise.
1361          */
1362         return result;
1363 }
1364
1365 /* ----------------------------------------------------------------
1366  *              ExecSelect
1367  *
1368  *              SELECTs are easy.. we just pass the tuple to the appropriate
1369  *              print function.  The only complexity is when we do a
1370  *              "SELECT INTO", in which case we insert the tuple into
1371  *              the appropriate relation (note: this is a newly created relation
1372  *              so we don't need to worry about indices or locks.)
1373  * ----------------------------------------------------------------
1374  */
1375 static void
1376 ExecSelect(TupleTableSlot *slot,
1377                    DestReceiver *dest,
1378                    EState *estate)
1379 {
1380         /*
1381          * insert the tuple into the "into relation"
1382          *
1383          * XXX this probably ought to be replaced by a separate destination
1384          */
1385         if (estate->es_into_relation_descriptor != NULL)
1386         {
1387                 HeapTuple       tuple;
1388
1389                 tuple = ExecCopySlotTuple(slot);
1390                 heap_insert(estate->es_into_relation_descriptor, tuple,
1391                                         estate->es_snapshot->curcid,
1392                                         estate->es_into_relation_use_wal,
1393                                         false);         /* never any point in using FSM */
1394                 /* we know there are no indexes to update */
1395                 heap_freetuple(tuple);
1396                 IncrAppended();
1397         }
1398
1399         /*
1400          * send the tuple to the destination
1401          */
1402         (*dest->receiveSlot) (slot, dest);
1403         IncrRetrieved();
1404         (estate->es_processed)++;
1405 }
1406
1407 /* ----------------------------------------------------------------
1408  *              ExecInsert
1409  *
1410  *              INSERTs are trickier.. we have to insert the tuple into
1411  *              the base relation and insert appropriate tuples into the
1412  *              index relations.
1413  * ----------------------------------------------------------------
1414  */
1415 static void
1416 ExecInsert(TupleTableSlot *slot,
1417                    ItemPointer tupleid,
1418                    EState *estate)
1419 {
1420         HeapTuple       tuple;
1421         ResultRelInfo *resultRelInfo;
1422         Relation        resultRelationDesc;
1423         Oid                     newId;
1424
1425         /*
1426          * get the heap tuple out of the tuple table slot, making sure we have a
1427          * writable copy
1428          */
1429         tuple = ExecMaterializeSlot(slot);
1430
1431         /*
1432          * get information on the (current) result relation
1433          */
1434         resultRelInfo = estate->es_result_relation_info;
1435         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1436
1437         /* BEFORE ROW INSERT Triggers */
1438         if (resultRelInfo->ri_TrigDesc &&
1439                 resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_INSERT] > 0)
1440         {
1441                 HeapTuple       newtuple;
1442
1443                 newtuple = ExecBRInsertTriggers(estate, resultRelInfo, tuple);
1444
1445                 if (newtuple == NULL)   /* "do nothing" */
1446                         return;
1447
1448                 if (newtuple != tuple)  /* modified by Trigger(s) */
1449                 {
1450                         /*
1451                          * Put the modified tuple into a slot for convenience of routines
1452                          * below.  We assume the tuple was allocated in per-tuple memory
1453                          * context, and therefore will go away by itself. The tuple table
1454                          * slot should not try to clear it.
1455                          */
1456                         TupleTableSlot *newslot = estate->es_trig_tuple_slot;
1457
1458                         if (newslot->tts_tupleDescriptor != slot->tts_tupleDescriptor)
1459                                 ExecSetSlotDescriptor(newslot, slot->tts_tupleDescriptor);
1460                         ExecStoreTuple(newtuple, newslot, InvalidBuffer, false);
1461                         slot = newslot;
1462                         tuple = newtuple;
1463                 }
1464         }
1465
1466         /*
1467          * Check the constraints of the tuple
1468          */
1469         if (resultRelationDesc->rd_att->constr)
1470                 ExecConstraints(resultRelInfo, slot, estate);
1471
1472         /*
1473          * insert the tuple
1474          *
1475          * Note: heap_insert returns the tid (location) of the new tuple in the
1476          * t_self field.
1477          */
1478         newId = heap_insert(resultRelationDesc, tuple,
1479                                                 estate->es_snapshot->curcid,
1480                                                 true, true);
1481
1482         IncrAppended();
1483         (estate->es_processed)++;
1484         estate->es_lastoid = newId;
1485         setLastTid(&(tuple->t_self));
1486
1487         /*
1488          * insert index entries for tuple
1489          */
1490         if (resultRelInfo->ri_NumIndices > 0)
1491                 ExecInsertIndexTuples(slot, &(tuple->t_self), estate, false);
1492
1493         /* AFTER ROW INSERT Triggers */
1494         ExecARInsertTriggers(estate, resultRelInfo, tuple);
1495 }
1496
1497 /* ----------------------------------------------------------------
1498  *              ExecDelete
1499  *
1500  *              DELETE is like UPDATE, except that we delete the tuple and no
1501  *              index modifications are needed
1502  * ----------------------------------------------------------------
1503  */
1504 static void
1505 ExecDelete(TupleTableSlot *slot,
1506                    ItemPointer tupleid,
1507                    EState *estate)
1508 {
1509         ResultRelInfo *resultRelInfo;
1510         Relation        resultRelationDesc;
1511         HTSU_Result result;
1512         ItemPointerData update_ctid;
1513         TransactionId update_xmax;
1514
1515         /*
1516          * get information on the (current) result relation
1517          */
1518         resultRelInfo = estate->es_result_relation_info;
1519         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1520
1521         /* BEFORE ROW DELETE Triggers */
1522         if (resultRelInfo->ri_TrigDesc &&
1523                 resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_DELETE] > 0)
1524         {
1525                 bool            dodelete;
1526
1527                 dodelete = ExecBRDeleteTriggers(estate, resultRelInfo, tupleid,
1528                                                                                 estate->es_snapshot->curcid);
1529
1530                 if (!dodelete)                  /* "do nothing" */
1531                         return;
1532         }
1533
1534         /*
1535          * delete the tuple
1536          *
1537          * Note: if es_crosscheck_snapshot isn't InvalidSnapshot, we check that
1538          * the row to be deleted is visible to that snapshot, and throw a can't-
1539          * serialize error if not.      This is a special-case behavior needed for
1540          * referential integrity updates in serializable transactions.
1541          */
1542 ldelete:;
1543         result = heap_delete(resultRelationDesc, tupleid,
1544                                                  &update_ctid, &update_xmax,
1545                                                  estate->es_snapshot->curcid,
1546                                                  estate->es_crosscheck_snapshot,
1547                                                  true /* wait for commit */ );
1548         switch (result)
1549         {
1550                 case HeapTupleSelfUpdated:
1551                         /* already deleted by self; nothing to do */
1552                         return;
1553
1554                 case HeapTupleMayBeUpdated:
1555                         break;
1556
1557                 case HeapTupleUpdated:
1558                         if (IsXactIsoLevelSerializable)
1559                                 ereport(ERROR,
1560                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1561                                                  errmsg("could not serialize access due to concurrent update")));
1562                         else if (!ItemPointerEquals(tupleid, &update_ctid))
1563                         {
1564                                 TupleTableSlot *epqslot;
1565
1566                                 epqslot = EvalPlanQual(estate,
1567                                                                            resultRelInfo->ri_RangeTableIndex,
1568                                                                            &update_ctid,
1569                                                                            update_xmax,
1570                                                                            estate->es_snapshot->curcid);
1571                                 if (!TupIsNull(epqslot))
1572                                 {
1573                                         *tupleid = update_ctid;
1574                                         goto ldelete;
1575                                 }
1576                         }
1577                         /* tuple already deleted; nothing to do */
1578                         return;
1579
1580                 default:
1581                         elog(ERROR, "unrecognized heap_delete status: %u", result);
1582                         return;
1583         }
1584
1585         IncrDeleted();
1586         (estate->es_processed)++;
1587
1588         /*
1589          * Note: Normally one would think that we have to delete index tuples
1590          * associated with the heap tuple now...
1591          *
1592          * ... but in POSTGRES, we have no need to do this because VACUUM will
1593          * take care of it later.  We can't delete index tuples immediately
1594          * anyway, since the tuple is still visible to other transactions.
1595          */
1596
1597         /* AFTER ROW DELETE Triggers */
1598         ExecARDeleteTriggers(estate, resultRelInfo, tupleid);
1599 }
1600
1601 /* ----------------------------------------------------------------
1602  *              ExecUpdate
1603  *
1604  *              note: we can't run UPDATE queries with transactions
1605  *              off because UPDATEs are actually INSERTs and our
1606  *              scan will mistakenly loop forever, updating the tuple
1607  *              it just inserted..      This should be fixed but until it
1608  *              is, we don't want to get stuck in an infinite loop
1609  *              which corrupts your database..
1610  * ----------------------------------------------------------------
1611  */
1612 static void
1613 ExecUpdate(TupleTableSlot *slot,
1614                    ItemPointer tupleid,
1615                    EState *estate)
1616 {
1617         HeapTuple       tuple;
1618         ResultRelInfo *resultRelInfo;
1619         Relation        resultRelationDesc;
1620         HTSU_Result result;
1621         ItemPointerData update_ctid;
1622         TransactionId update_xmax;
1623
1624         /*
1625          * abort the operation if not running transactions
1626          */
1627         if (IsBootstrapProcessingMode())
1628                 elog(ERROR, "cannot UPDATE during bootstrap");
1629
1630         /*
1631          * get the heap tuple out of the tuple table slot, making sure we have a
1632          * writable copy
1633          */
1634         tuple = ExecMaterializeSlot(slot);
1635
1636         /*
1637          * get information on the (current) result relation
1638          */
1639         resultRelInfo = estate->es_result_relation_info;
1640         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1641
1642         /* BEFORE ROW UPDATE Triggers */
1643         if (resultRelInfo->ri_TrigDesc &&
1644                 resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_UPDATE] > 0)
1645         {
1646                 HeapTuple       newtuple;
1647
1648                 newtuple = ExecBRUpdateTriggers(estate, resultRelInfo,
1649                                                                                 tupleid, tuple,
1650                                                                                 estate->es_snapshot->curcid);
1651
1652                 if (newtuple == NULL)   /* "do nothing" */
1653                         return;
1654
1655                 if (newtuple != tuple)  /* modified by Trigger(s) */
1656                 {
1657                         /*
1658                          * Put the modified tuple into a slot for convenience of routines
1659                          * below.  We assume the tuple was allocated in per-tuple memory
1660                          * context, and therefore will go away by itself. The tuple table
1661                          * slot should not try to clear it.
1662                          */
1663                         TupleTableSlot *newslot = estate->es_trig_tuple_slot;
1664
1665                         if (newslot->tts_tupleDescriptor != slot->tts_tupleDescriptor)
1666                                 ExecSetSlotDescriptor(newslot, slot->tts_tupleDescriptor);
1667                         ExecStoreTuple(newtuple, newslot, InvalidBuffer, false);
1668                         slot = newslot;
1669                         tuple = newtuple;
1670                 }
1671         }
1672
1673         /*
1674          * Check the constraints of the tuple
1675          *
1676          * If we generate a new candidate tuple after EvalPlanQual testing, we
1677          * must loop back here and recheck constraints.  (We don't need to redo
1678          * triggers, however.  If there are any BEFORE triggers then trigger.c
1679          * will have done heap_lock_tuple to lock the correct tuple, so there's no
1680          * need to do them again.)
1681          */
1682 lreplace:;
1683         if (resultRelationDesc->rd_att->constr)
1684                 ExecConstraints(resultRelInfo, slot, estate);
1685
1686         /*
1687          * replace the heap tuple
1688          *
1689          * Note: if es_crosscheck_snapshot isn't InvalidSnapshot, we check that
1690          * the row to be updated is visible to that snapshot, and throw a can't-
1691          * serialize error if not.      This is a special-case behavior needed for
1692          * referential integrity updates in serializable transactions.
1693          */
1694         result = heap_update(resultRelationDesc, tupleid, tuple,
1695                                                  &update_ctid, &update_xmax,
1696                                                  estate->es_snapshot->curcid,
1697                                                  estate->es_crosscheck_snapshot,
1698                                                  true /* wait for commit */ );
1699         switch (result)
1700         {
1701                 case HeapTupleSelfUpdated:
1702                         /* already deleted by self; nothing to do */
1703                         return;
1704
1705                 case HeapTupleMayBeUpdated:
1706                         break;
1707
1708                 case HeapTupleUpdated:
1709                         if (IsXactIsoLevelSerializable)
1710                                 ereport(ERROR,
1711                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1712                                                  errmsg("could not serialize access due to concurrent update")));
1713                         else if (!ItemPointerEquals(tupleid, &update_ctid))
1714                         {
1715                                 TupleTableSlot *epqslot;
1716
1717                                 epqslot = EvalPlanQual(estate,
1718                                                                            resultRelInfo->ri_RangeTableIndex,
1719                                                                            &update_ctid,
1720                                                                            update_xmax,
1721                                                                            estate->es_snapshot->curcid);
1722                                 if (!TupIsNull(epqslot))
1723                                 {
1724                                         *tupleid = update_ctid;
1725                                         slot = ExecFilterJunk(estate->es_junkFilter, epqslot);
1726                                         tuple = ExecMaterializeSlot(slot);
1727                                         goto lreplace;
1728                                 }
1729                         }
1730                         /* tuple already deleted; nothing to do */
1731                         return;
1732
1733                 default:
1734                         elog(ERROR, "unrecognized heap_update status: %u", result);
1735                         return;
1736         }
1737
1738         IncrReplaced();
1739         (estate->es_processed)++;
1740
1741         /*
1742          * Note: instead of having to update the old index tuples associated with
1743          * the heap tuple, all we do is form and insert new index tuples. This is
1744          * because UPDATEs are actually DELETEs and INSERTs, and index tuple
1745          * deletion is done later by VACUUM (see notes in ExecDelete).  All we do
1746          * here is insert new index tuples.  -cim 9/27/89
1747          */
1748
1749         /*
1750          * insert index entries for tuple
1751          *
1752          * Note: heap_update returns the tid (location) of the new tuple in the
1753          * t_self field.
1754          */
1755         if (resultRelInfo->ri_NumIndices > 0)
1756                 ExecInsertIndexTuples(slot, &(tuple->t_self), estate, false);
1757
1758         /* AFTER ROW UPDATE Triggers */
1759         ExecARUpdateTriggers(estate, resultRelInfo, tupleid, tuple);
1760 }
1761
1762 static const char *
1763 ExecRelCheck(ResultRelInfo *resultRelInfo,
1764                          TupleTableSlot *slot, EState *estate)
1765 {
1766         Relation        rel = resultRelInfo->ri_RelationDesc;
1767         int                     ncheck = rel->rd_att->constr->num_check;
1768         ConstrCheck *check = rel->rd_att->constr->check;
1769         ExprContext *econtext;
1770         MemoryContext oldContext;
1771         List       *qual;
1772         int                     i;
1773
1774         /*
1775          * If first time through for this result relation, build expression
1776          * nodetrees for rel's constraint expressions.  Keep them in the per-query
1777          * memory context so they'll survive throughout the query.
1778          */
1779         if (resultRelInfo->ri_ConstraintExprs == NULL)
1780         {
1781                 oldContext = MemoryContextSwitchTo(estate->es_query_cxt);
1782                 resultRelInfo->ri_ConstraintExprs =
1783                         (List **) palloc(ncheck * sizeof(List *));
1784                 for (i = 0; i < ncheck; i++)
1785                 {
1786                         /* ExecQual wants implicit-AND form */
1787                         qual = make_ands_implicit(stringToNode(check[i].ccbin));
1788                         resultRelInfo->ri_ConstraintExprs[i] = (List *)
1789                                 ExecPrepareExpr((Expr *) qual, estate);
1790                 }
1791                 MemoryContextSwitchTo(oldContext);
1792         }
1793
1794         /*
1795          * We will use the EState's per-tuple context for evaluating constraint
1796          * expressions (creating it if it's not already there).
1797          */
1798         econtext = GetPerTupleExprContext(estate);
1799
1800         /* Arrange for econtext's scan tuple to be the tuple under test */
1801         econtext->ecxt_scantuple = slot;
1802
1803         /* And evaluate the constraints */
1804         for (i = 0; i < ncheck; i++)
1805         {
1806                 qual = resultRelInfo->ri_ConstraintExprs[i];
1807
1808                 /*
1809                  * NOTE: SQL92 specifies that a NULL result from a constraint
1810                  * expression is not to be treated as a failure.  Therefore, tell
1811                  * ExecQual to return TRUE for NULL.
1812                  */
1813                 if (!ExecQual(qual, econtext, true))
1814                         return check[i].ccname;
1815         }
1816
1817         /* NULL result means no error */
1818         return NULL;
1819 }
1820
1821 void
1822 ExecConstraints(ResultRelInfo *resultRelInfo,
1823                                 TupleTableSlot *slot, EState *estate)
1824 {
1825         Relation        rel = resultRelInfo->ri_RelationDesc;
1826         TupleConstr *constr = rel->rd_att->constr;
1827
1828         Assert(constr);
1829
1830         if (constr->has_not_null)
1831         {
1832                 int                     natts = rel->rd_att->natts;
1833                 int                     attrChk;
1834
1835                 for (attrChk = 1; attrChk <= natts; attrChk++)
1836                 {
1837                         if (rel->rd_att->attrs[attrChk - 1]->attnotnull &&
1838                                 slot_attisnull(slot, attrChk))
1839                                 ereport(ERROR,
1840                                                 (errcode(ERRCODE_NOT_NULL_VIOLATION),
1841                                                  errmsg("null value in column \"%s\" violates not-null constraint",
1842                                                 NameStr(rel->rd_att->attrs[attrChk - 1]->attname))));
1843                 }
1844         }
1845
1846         if (constr->num_check > 0)
1847         {
1848                 const char *failed;
1849
1850                 if ((failed = ExecRelCheck(resultRelInfo, slot, estate)) != NULL)
1851                         ereport(ERROR,
1852                                         (errcode(ERRCODE_CHECK_VIOLATION),
1853                                          errmsg("new row for relation \"%s\" violates check constraint \"%s\"",
1854                                                         RelationGetRelationName(rel), failed)));
1855         }
1856 }
1857
1858 /*
1859  * Check a modified tuple to see if we want to process its updated version
1860  * under READ COMMITTED rules.
1861  *
1862  * See backend/executor/README for some info about how this works.
1863  *
1864  *      estate - executor state data
1865  *      rti - rangetable index of table containing tuple
1866  *      *tid - t_ctid from the outdated tuple (ie, next updated version)
1867  *      priorXmax - t_xmax from the outdated tuple
1868  *      curCid - command ID of current command of my transaction
1869  *
1870  * *tid is also an output parameter: it's modified to hold the TID of the
1871  * latest version of the tuple (note this may be changed even on failure)
1872  *
1873  * Returns a slot containing the new candidate update/delete tuple, or
1874  * NULL if we determine we shouldn't process the row.
1875  */
1876 TupleTableSlot *
1877 EvalPlanQual(EState *estate, Index rti,
1878                          ItemPointer tid, TransactionId priorXmax, CommandId curCid)
1879 {
1880         evalPlanQual *epq;
1881         EState     *epqstate;
1882         Relation        relation;
1883         HeapTupleData tuple;
1884         HeapTuple       copyTuple = NULL;
1885         bool            endNode;
1886
1887         Assert(rti != 0);
1888
1889         /*
1890          * find relation containing target tuple
1891          */
1892         if (estate->es_result_relation_info != NULL &&
1893                 estate->es_result_relation_info->ri_RangeTableIndex == rti)
1894                 relation = estate->es_result_relation_info->ri_RelationDesc;
1895         else
1896         {
1897                 ListCell   *l;
1898
1899                 relation = NULL;
1900                 foreach(l, estate->es_rowMarks)
1901                 {
1902                         if (((ExecRowMark *) lfirst(l))->rti == rti)
1903                         {
1904                                 relation = ((ExecRowMark *) lfirst(l))->relation;
1905                                 break;
1906                         }
1907                 }
1908                 if (relation == NULL)
1909                         elog(ERROR, "could not find RowMark for RT index %u", rti);
1910         }
1911
1912         /*
1913          * fetch tid tuple
1914          *
1915          * Loop here to deal with updated or busy tuples
1916          */
1917         tuple.t_self = *tid;
1918         for (;;)
1919         {
1920                 Buffer          buffer;
1921
1922                 if (heap_fetch(relation, SnapshotDirty, &tuple, &buffer, true, NULL))
1923                 {
1924                         /*
1925                          * If xmin isn't what we're expecting, the slot must have been
1926                          * recycled and reused for an unrelated tuple.  This implies that
1927                          * the latest version of the row was deleted, so we need do
1928                          * nothing.  (Should be safe to examine xmin without getting
1929                          * buffer's content lock, since xmin never changes in an existing
1930                          * tuple.)
1931                          */
1932                         if (!TransactionIdEquals(HeapTupleHeaderGetXmin(tuple.t_data),
1933                                                                          priorXmax))
1934                         {
1935                                 ReleaseBuffer(buffer);
1936                                 return NULL;
1937                         }
1938
1939                         /* otherwise xmin should not be dirty... */
1940                         if (TransactionIdIsValid(SnapshotDirty->xmin))
1941                                 elog(ERROR, "t_xmin is uncommitted in tuple to be updated");
1942
1943                         /*
1944                          * If tuple is being updated by other transaction then we have to
1945                          * wait for its commit/abort.
1946                          */
1947                         if (TransactionIdIsValid(SnapshotDirty->xmax))
1948                         {
1949                                 ReleaseBuffer(buffer);
1950                                 XactLockTableWait(SnapshotDirty->xmax);
1951                                 continue;               /* loop back to repeat heap_fetch */
1952                         }
1953
1954                         /*
1955                          * If tuple was inserted by our own transaction, we have to check
1956                          * cmin against curCid: cmin >= curCid means our command cannot
1957                          * see the tuple, so we should ignore it.  Without this we are
1958                          * open to the "Halloween problem" of indefinitely re-updating
1959                          * the same tuple.  (We need not check cmax because
1960                          * HeapTupleSatisfiesDirty will consider a tuple deleted by
1961                          * our transaction dead, regardless of cmax.)  We just checked
1962                          * that priorXmax == xmin, so we can test that variable instead
1963                          * of doing HeapTupleHeaderGetXmin again.
1964                          */
1965                         if (TransactionIdIsCurrentTransactionId(priorXmax) &&
1966                                 HeapTupleHeaderGetCmin(tuple.t_data) >= curCid)
1967                         {
1968                                 ReleaseBuffer(buffer);
1969                                 return NULL;
1970                         }
1971
1972                         /*
1973                          * We got tuple - now copy it for use by recheck query.
1974                          */
1975                         copyTuple = heap_copytuple(&tuple);
1976                         ReleaseBuffer(buffer);
1977                         break;
1978                 }
1979
1980                 /*
1981                  * If the referenced slot was actually empty, the latest version of
1982                  * the row must have been deleted, so we need do nothing.
1983                  */
1984                 if (tuple.t_data == NULL)
1985                 {
1986                         ReleaseBuffer(buffer);
1987                         return NULL;
1988                 }
1989
1990                 /*
1991                  * As above, if xmin isn't what we're expecting, do nothing.
1992                  */
1993                 if (!TransactionIdEquals(HeapTupleHeaderGetXmin(tuple.t_data),
1994                                                                  priorXmax))
1995                 {
1996                         ReleaseBuffer(buffer);
1997                         return NULL;
1998                 }
1999
2000                 /*
2001                  * If we get here, the tuple was found but failed SnapshotDirty.
2002                  * Assuming the xmin is either a committed xact or our own xact (as it
2003                  * certainly should be if we're trying to modify the tuple), this must
2004                  * mean that the row was updated or deleted by either a committed xact
2005                  * or our own xact.  If it was deleted, we can ignore it; if it was
2006                  * updated then chain up to the next version and repeat the whole
2007                  * test.
2008                  *
2009                  * As above, it should be safe to examine xmax and t_ctid without the
2010                  * buffer content lock, because they can't be changing.
2011                  */
2012                 if (ItemPointerEquals(&tuple.t_self, &tuple.t_data->t_ctid))
2013                 {
2014                         /* deleted, so forget about it */
2015                         ReleaseBuffer(buffer);
2016                         return NULL;
2017                 }
2018
2019                 /* updated, so look at the updated row */
2020                 tuple.t_self = tuple.t_data->t_ctid;
2021                 /* updated row should have xmin matching this xmax */
2022                 priorXmax = HeapTupleHeaderGetXmax(tuple.t_data);
2023                 ReleaseBuffer(buffer);
2024                 /* loop back to fetch next in chain */
2025         }
2026
2027         /*
2028          * For UPDATE/DELETE we have to return tid of actual row we're executing
2029          * PQ for.
2030          */
2031         *tid = tuple.t_self;
2032
2033         /*
2034          * Need to run a recheck subquery.      Find or create a PQ stack entry.
2035          */
2036         epq = estate->es_evalPlanQual;
2037         endNode = true;
2038
2039         if (epq != NULL && epq->rti == 0)
2040         {
2041                 /* Top PQ stack entry is idle, so re-use it */
2042                 Assert(!(estate->es_useEvalPlan) && epq->next == NULL);
2043                 epq->rti = rti;
2044                 endNode = false;
2045         }
2046
2047         /*
2048          * If this is request for another RTE - Ra, - then we have to check wasn't
2049          * PlanQual requested for Ra already and if so then Ra' row was updated
2050          * again and we have to re-start old execution for Ra and forget all what
2051          * we done after Ra was suspended. Cool? -:))
2052          */
2053         if (epq != NULL && epq->rti != rti &&
2054                 epq->estate->es_evTuple[rti - 1] != NULL)
2055         {
2056                 do
2057                 {
2058                         evalPlanQual *oldepq;
2059
2060                         /* stop execution */
2061                         EvalPlanQualStop(epq);
2062                         /* pop previous PlanQual from the stack */
2063                         oldepq = epq->next;
2064                         Assert(oldepq && oldepq->rti != 0);
2065                         /* push current PQ to freePQ stack */
2066                         oldepq->free = epq;
2067                         epq = oldepq;
2068                         estate->es_evalPlanQual = epq;
2069                 } while (epq->rti != rti);
2070         }
2071
2072         /*
2073          * If we are requested for another RTE then we have to suspend execution
2074          * of current PlanQual and start execution for new one.
2075          */
2076         if (epq == NULL || epq->rti != rti)
2077         {
2078                 /* try to reuse plan used previously */
2079                 evalPlanQual *newepq = (epq != NULL) ? epq->free : NULL;
2080
2081                 if (newepq == NULL)             /* first call or freePQ stack is empty */
2082                 {
2083                         newepq = (evalPlanQual *) palloc0(sizeof(evalPlanQual));
2084                         newepq->free = NULL;
2085                         newepq->estate = NULL;
2086                         newepq->planstate = NULL;
2087                 }
2088                 else
2089                 {
2090                         /* recycle previously used PlanQual */
2091                         Assert(newepq->estate == NULL);
2092                         epq->free = NULL;
2093                 }
2094                 /* push current PQ to the stack */
2095                 newepq->next = epq;
2096                 epq = newepq;
2097                 estate->es_evalPlanQual = epq;
2098                 epq->rti = rti;
2099                 endNode = false;
2100         }
2101
2102         Assert(epq->rti == rti);
2103
2104         /*
2105          * Ok - we're requested for the same RTE.  Unfortunately we still have to
2106          * end and restart execution of the plan, because ExecReScan wouldn't
2107          * ensure that upper plan nodes would reset themselves.  We could make
2108          * that work if insertion of the target tuple were integrated with the
2109          * Param mechanism somehow, so that the upper plan nodes know that their
2110          * children's outputs have changed.
2111          *
2112          * Note that the stack of free evalPlanQual nodes is quite useless at the
2113          * moment, since it only saves us from pallocing/releasing the
2114          * evalPlanQual nodes themselves.  But it will be useful once we implement
2115          * ReScan instead of end/restart for re-using PlanQual nodes.
2116          */
2117         if (endNode)
2118         {
2119                 /* stop execution */
2120                 EvalPlanQualStop(epq);
2121         }
2122
2123         /*
2124          * Initialize new recheck query.
2125          *
2126          * Note: if we were re-using PlanQual plans via ExecReScan, we'd need to
2127          * instead copy down changeable state from the top plan (including
2128          * es_result_relation_info, es_junkFilter) and reset locally changeable
2129          * state in the epq (including es_param_exec_vals, es_evTupleNull).
2130          */
2131         EvalPlanQualStart(epq, estate, epq->next);
2132
2133         /*
2134          * free old RTE' tuple, if any, and store target tuple where relation's
2135          * scan node will see it
2136          */
2137         epqstate = epq->estate;
2138         if (epqstate->es_evTuple[rti - 1] != NULL)
2139                 heap_freetuple(epqstate->es_evTuple[rti - 1]);
2140         epqstate->es_evTuple[rti - 1] = copyTuple;
2141
2142         return EvalPlanQualNext(estate);
2143 }
2144
2145 static TupleTableSlot *
2146 EvalPlanQualNext(EState *estate)
2147 {
2148         evalPlanQual *epq = estate->es_evalPlanQual;
2149         MemoryContext oldcontext;
2150         TupleTableSlot *slot;
2151
2152         Assert(epq->rti != 0);
2153
2154 lpqnext:;
2155         oldcontext = MemoryContextSwitchTo(epq->estate->es_query_cxt);
2156         slot = ExecProcNode(epq->planstate);
2157         MemoryContextSwitchTo(oldcontext);
2158
2159         /*
2160          * No more tuples for this PQ. Continue previous one.
2161          */
2162         if (TupIsNull(slot))
2163         {
2164                 evalPlanQual *oldepq;
2165
2166                 /* stop execution */
2167                 EvalPlanQualStop(epq);
2168                 /* pop old PQ from the stack */
2169                 oldepq = epq->next;
2170                 if (oldepq == NULL)
2171                 {
2172                         /* this is the first (oldest) PQ - mark as free */
2173                         epq->rti = 0;
2174                         estate->es_useEvalPlan = false;
2175                         /* and continue Query execution */
2176                         return NULL;
2177                 }
2178                 Assert(oldepq->rti != 0);
2179                 /* push current PQ to freePQ stack */
2180                 oldepq->free = epq;
2181                 epq = oldepq;
2182                 estate->es_evalPlanQual = epq;
2183                 goto lpqnext;
2184         }
2185
2186         return slot;
2187 }
2188
2189 static void
2190 EndEvalPlanQual(EState *estate)
2191 {
2192         evalPlanQual *epq = estate->es_evalPlanQual;
2193
2194         if (epq->rti == 0)                      /* plans already shutdowned */
2195         {
2196                 Assert(epq->next == NULL);
2197                 return;
2198         }
2199
2200         for (;;)
2201         {
2202                 evalPlanQual *oldepq;
2203
2204                 /* stop execution */
2205                 EvalPlanQualStop(epq);
2206                 /* pop old PQ from the stack */
2207                 oldepq = epq->next;
2208                 if (oldepq == NULL)
2209                 {
2210                         /* this is the first (oldest) PQ - mark as free */
2211                         epq->rti = 0;
2212                         estate->es_useEvalPlan = false;
2213                         break;
2214                 }
2215                 Assert(oldepq->rti != 0);
2216                 /* push current PQ to freePQ stack */
2217                 oldepq->free = epq;
2218                 epq = oldepq;
2219                 estate->es_evalPlanQual = epq;
2220         }
2221 }
2222
2223 /*
2224  * Start execution of one level of PlanQual.
2225  *
2226  * This is a cut-down version of ExecutorStart(): we copy some state from
2227  * the top-level estate rather than initializing it fresh.
2228  */
2229 static void
2230 EvalPlanQualStart(evalPlanQual *epq, EState *estate, evalPlanQual *priorepq)
2231 {
2232         EState     *epqstate;
2233         int                     rtsize;
2234         MemoryContext oldcontext;
2235
2236         rtsize = list_length(estate->es_range_table);
2237
2238         epq->estate = epqstate = CreateExecutorState();
2239
2240         oldcontext = MemoryContextSwitchTo(epqstate->es_query_cxt);
2241
2242         /*
2243          * The epqstates share the top query's copy of unchanging state such as
2244          * the snapshot, rangetable, result-rel info, and external Param info.
2245          * They need their own copies of local state, including a tuple table,
2246          * es_param_exec_vals, etc.
2247          */
2248         epqstate->es_direction = ForwardScanDirection;
2249         epqstate->es_snapshot = estate->es_snapshot;
2250         epqstate->es_crosscheck_snapshot = estate->es_crosscheck_snapshot;
2251         epqstate->es_range_table = estate->es_range_table;
2252         epqstate->es_result_relations = estate->es_result_relations;
2253         epqstate->es_num_result_relations = estate->es_num_result_relations;
2254         epqstate->es_result_relation_info = estate->es_result_relation_info;
2255         epqstate->es_junkFilter = estate->es_junkFilter;
2256         epqstate->es_into_relation_descriptor = estate->es_into_relation_descriptor;
2257         epqstate->es_into_relation_use_wal = estate->es_into_relation_use_wal;
2258         epqstate->es_param_list_info = estate->es_param_list_info;
2259         if (estate->es_topPlan->nParamExec > 0)
2260                 epqstate->es_param_exec_vals = (ParamExecData *)
2261                         palloc0(estate->es_topPlan->nParamExec * sizeof(ParamExecData));
2262         epqstate->es_rowMarks = estate->es_rowMarks;
2263         epqstate->es_instrument = estate->es_instrument;
2264         epqstate->es_select_into = estate->es_select_into;
2265         epqstate->es_into_oids = estate->es_into_oids;
2266         epqstate->es_topPlan = estate->es_topPlan;
2267
2268         /*
2269          * Each epqstate must have its own es_evTupleNull state, but all the stack
2270          * entries share es_evTuple state.      This allows sub-rechecks to inherit
2271          * the value being examined by an outer recheck.
2272          */
2273         epqstate->es_evTupleNull = (bool *) palloc0(rtsize * sizeof(bool));
2274         if (priorepq == NULL)
2275                 /* first PQ stack entry */
2276                 epqstate->es_evTuple = (HeapTuple *)
2277                         palloc0(rtsize * sizeof(HeapTuple));
2278         else
2279                 /* later stack entries share the same storage */
2280                 epqstate->es_evTuple = priorepq->estate->es_evTuple;
2281
2282         epqstate->es_tupleTable =
2283                 ExecCreateTupleTable(estate->es_tupleTable->size);
2284
2285         epq->planstate = ExecInitNode(estate->es_topPlan, epqstate, 0);
2286
2287         MemoryContextSwitchTo(oldcontext);
2288 }
2289
2290 /*
2291  * End execution of one level of PlanQual.
2292  *
2293  * This is a cut-down version of ExecutorEnd(); basically we want to do most
2294  * of the normal cleanup, but *not* close result relations (which we are
2295  * just sharing from the outer query).
2296  */
2297 static void
2298 EvalPlanQualStop(evalPlanQual *epq)
2299 {
2300         EState     *epqstate = epq->estate;
2301         MemoryContext oldcontext;
2302
2303         oldcontext = MemoryContextSwitchTo(epqstate->es_query_cxt);
2304
2305         ExecEndNode(epq->planstate);
2306
2307         ExecDropTupleTable(epqstate->es_tupleTable, true);
2308         epqstate->es_tupleTable = NULL;
2309
2310         if (epqstate->es_evTuple[epq->rti - 1] != NULL)
2311         {
2312                 heap_freetuple(epqstate->es_evTuple[epq->rti - 1]);
2313                 epqstate->es_evTuple[epq->rti - 1] = NULL;
2314         }
2315
2316         MemoryContextSwitchTo(oldcontext);
2317
2318         FreeExecutorState(epqstate);
2319
2320         epq->estate = NULL;
2321         epq->planstate = NULL;
2322 }